From dfe08f395fb7860e5b888f37231ad8564852b95a Mon Sep 17 00:00:00 2001
From: BootsofLagrangian <125134079+BootsofLagrangian@users.noreply.github.com>
Date: Sun, 4 Feb 2024 03:12:42 +0900
Subject: [PATCH 01/69] support deepspeed

---
 fine_tune.py          | 41 ++++++++++++++++++++-------
 library/train_util.py | 54 +++++++++++++++++++++++++++++++++++
 sdxl_train.py         | 45 ++++++++++++++++++++---------
 train_db.py           | 37 ++++++++++++++++++------
 train_network.py      | 66 +++++++++++++++++++++++++++++++------------
 5 files changed, 194 insertions(+), 49 deletions(-)

diff --git a/fine_tune.py b/fine_tune.py
index 982dc8aec..78dfd1696 100644
--- a/fine_tune.py
+++ b/fine_tune.py
@@ -102,6 +102,7 @@ def train(args):
 
     # mixed precisionに対応した型を用意しておき適宜castする
     weight_dtype, save_dtype = train_util.prepare_dtype(args)
+    vae_dtype = torch.float32 if args.no_half_vae else weight_dtype
 
     # モデルを読み込む
     text_encoder, vae, unet, load_stable_diffusion_format = train_util.load_target_model(args, weight_dtype, accelerator)
@@ -152,7 +153,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
 
     # 学習を準備する
     if cache_latents:
-        vae.to(accelerator.device, dtype=weight_dtype)
+        vae.to(accelerator.device, dtype=vae_dtype)
         vae.requires_grad_(False)
         vae.eval()
         with torch.no_grad():
@@ -187,7 +188,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
     if not cache_latents:
         vae.requires_grad_(False)
         vae.eval()
-        vae.to(accelerator.device, dtype=weight_dtype)
+        vae.to(accelerator.device, dtype=vae_dtype)
 
     for m in training_models:
         m.requires_grad_(True)
@@ -214,7 +215,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
         batch_size=1,
         shuffle=True,
         collate_fn=collator,
-        num_workers=n_workers,
+        num_workers=n_workers if not args.deepspeed else 1, # To avoid RuntimeError: DataLoader worker exited unexpectedly with exit code 1.
         persistent_workers=args.persistent_data_loader_workers,
     )
 
@@ -240,13 +241,33 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
         unet.to(weight_dtype)
         text_encoder.to(weight_dtype)
 
-    # acceleratorがなんかよろしくやってくれるらしい
-    if args.train_text_encoder:
-        unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-            unet, text_encoder, optimizer, train_dataloader, lr_scheduler
-        )
-    else:
-        unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(unet, optimizer, train_dataloader, lr_scheduler)
+    if args.deepspeed:
+        # wrapping model
+        class DeepSpeedModel(torch.nn.Module): 
+            def __init__(self, unet, text_encoder, vae) -> None:
+                super().__init__()
+                self.unet = unet
+                self.text_encoders = self.text_encoder = torch.nn.ModuleList(text_encoder)
+                self.vae = vae
+            def get_models(self):
+                return self.unet, self.text_encoders, self.vae
+            
+        unet.to(accelerator.device, dtype=weight_dtype)
+        [t_enc.to(accelerator.device, dtype=weight_dtype) for t_enc in text_encoders]
+        ds_model = DeepSpeedModel(unet, text_encoders, vae)
+        ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(ds_model, optimizer, train_dataloader, lr_scheduler)
+        # Now, ds_model is an instance of DeepSpeedEngine. 
+        unet, text_encoders, vae = ds_model.get_models() # for compatiblility
+        vae.to(vae_dtype)
+        text_encoder = text_encoders
+            
+    else: # acceleratorがなんかよろしくやってくれるらしい
+        if args.train_text_encoder:
+            unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+                unet, text_encoder, optimizer, train_dataloader, lr_scheduler
+            )
+        else:
+            unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(unet, optimizer, train_dataloader, lr_scheduler)
 
     # 実験的機能：勾配も含めたfp16学習を行う　PyTorchにパッチを当ててfp16でのgrad scaleを有効にする
     if args.full_fp16:
diff --git a/library/train_util.py b/library/train_util.py
index ba428e508..2d85c9776 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -20,6 +20,7 @@
     Union,
 )
 from accelerate import Accelerator, InitProcessGroupKwargs, DistributedDataParallelKwargs
+from accelerate import DeepSpeedPlugin
 import gc
 import glob
 import math
@@ -3124,6 +3125,47 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
             "--prior_loss_weight", type=float, default=1.0, help="loss weight for regularization images / 正則化画像のlossの重み"
         )
 
+    # DeepSpeed Arguments. https://huggingface.co/docs/accelerate/usage_guides/deepspeed
+    parser.add_argument("--deepspeed", action="store_true", help="enable deepspeed training")
+    parser.add_argument(
+        "--zero_stage", 
+        type=int, default=2,
+        choices=[0, 1, 2, 3],
+        help="Possible options are 0,1,2,3."
+    )
+    parser.add_argument(
+        "--offload_optimizer", 
+        type=str, default=None,
+        choices=[None, "cpu", "nvme"],
+        help="Possible options are none|cpu|nvme. Only applicable with ZeRO Stages 2 and 3."
+    )
+    parser.add_argument(
+        "--offload_optimizer_nvme_path",
+        type=str, default=None,
+        help="Possible options are /nvme|/local_nvme. Only applicable with ZeRO Stage 3."
+    )
+    parser.add_argument(
+        "--offload_param_device",
+        type=str, default=None,
+        choices=[None, "cpu", "nvme"],
+        help="Possible options are none|cpu|nvme. Only applicable with ZeRO Stage 3."
+    )
+    parser.add_argument(
+        "--offload_param_nvme_path",
+        type=str, default=None,
+        help="Possible options are /nvme|/local_nvme. Only applicable with ZeRO Stage 3."
+    )
+    parser.add_argument(
+        "--zero3_init_flag",
+        action="store_true",
+        help="Flag to indicate whether to enable `deepspeed.zero.Init` for constructing massive models."
+            "Only applicable with ZeRO Stage-3."
+    )
+    parser.add_argument(
+        "--zero3_save_16bit_model",
+        action="store_true",
+        help="Flag to indicate whether to save 16-bit model. Only applicable with ZeRO Stage-3."
+    )
 
 def verify_training_args(args: argparse.Namespace):
     if args.v_parameterization and not args.v2:
@@ -3912,6 +3954,17 @@ def prepare_accelerator(args: argparse.Namespace):
         else None,
     )
     kwargs_handlers = list(filter(lambda x: x is not None, kwargs_handlers))
+    deepspeed_plugin = None
+    if args.deepspeed:
+        deepspeed_plugin = DeepSpeedPlugin(
+            zero_stage=args.zero_stage,
+            gradient_accumulation_steps=args.gradient_accumulation_steps, gradient_clipping=args.max_grad_norm,
+            offload_optimizer=args.offload_optimizer, offload_optimizer_nvme_path=args.offload_optimizer_nvme_path,
+            offload_param_device=args.offload_param_device, offload_param_nvme_path=args.offload_param_nvme_path,
+            zero3_init_flag=args.zero3_init_flag, zero3_save_16bit_model=args.zero3_save_16bit_model,
+        )
+        deepspeed_plugin.deepspeed_config['train_micro_batch_size_per_gpu'] = args.train_batch_size
+
     accelerator = Accelerator(
         gradient_accumulation_steps=args.gradient_accumulation_steps,
         mixed_precision=args.mixed_precision,
@@ -3919,6 +3972,7 @@ def prepare_accelerator(args: argparse.Namespace):
         project_dir=logging_dir,
         kwargs_handlers=kwargs_handlers,
         dynamo_backend=dynamo_backend,
+        deepspeed_plugin=deepspeed_plugin,
     )
     return accelerator
 
diff --git a/sdxl_train.py b/sdxl_train.py
index a3f6f3a17..6ce6c201e 100644
--- a/sdxl_train.py
+++ b/sdxl_train.py
@@ -354,7 +354,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
         batch_size=1,
         shuffle=True,
         collate_fn=collator,
-        num_workers=n_workers,
+        num_workers=n_workers if not args.deepspeed else 1, # To avoid RuntimeError: DataLoader worker exited unexpectedly with exit code 1.
         persistent_workers=args.persistent_data_loader_workers,
     )
 
@@ -389,18 +389,37 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
         text_encoder1.to(weight_dtype)
         text_encoder2.to(weight_dtype)
 
-    # acceleratorがなんかよろしくやってくれるらしい
-    if train_unet:
-        unet = accelerator.prepare(unet)
-    if train_text_encoder1:
-        # freeze last layer and final_layer_norm in te1 since we use the output of the penultimate layer
-        text_encoder1.text_model.encoder.layers[-1].requires_grad_(False)
-        text_encoder1.text_model.final_layer_norm.requires_grad_(False)
-        text_encoder1 = accelerator.prepare(text_encoder1)
-    if train_text_encoder2:
-        text_encoder2 = accelerator.prepare(text_encoder2)
-
-    optimizer, train_dataloader, lr_scheduler = accelerator.prepare(optimizer, train_dataloader, lr_scheduler)
+    if args.deepspeed:
+        # Wrapping model for DeepSpeed
+        class DeepSpeedModel(torch.nn.Module): 
+            def __init__(self, unet, text_encoder, vae) -> None:
+                super().__init__()
+                self.unet = unet
+                self.text_encoders = self.text_encoder = torch.nn.ModuleList(text_encoder)
+                self.vae = vae
+                
+            def get_models(self):
+                return self.unet, self.text_encoders, self.vae
+        text_encoders = [text_encoder1, text_encoder2]
+        unet.to(accelerator.device, dtype=weight_dtype)
+        [t_enc.to(accelerator.device, dtype=weight_dtype) for t_enc in text_encoders]
+        ds_model = DeepSpeedModel(unet, text_encoders, vae)
+        ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(ds_model, optimizer, train_dataloader, lr_scheduler)
+        # Now, ds_model is an instance of DeepSpeedEngine. 
+        unet, text_encoders, vae = ds_model.get_models() # for compatiblility
+        vae.to(vae_dtype) # to avoid explicitly half-vae
+        text_encoder1, text_encoder2 = text_encoders[0], text_encoders[1]
+    else: # acceleratorがなんかよろしくやってくれるらしい
+        if train_unet:
+            unet = accelerator.prepare(unet)
+        if train_text_encoder1:
+            # freeze last layer and final_layer_norm in te1 since we use the output of the penultimate layer
+            text_encoder1.text_model.encoder.layers[-1].requires_grad_(False)
+            text_encoder1.text_model.final_layer_norm.requires_grad_(False)
+            text_encoder1 = accelerator.prepare(text_encoder1)
+        if train_text_encoder2:
+            text_encoder2 = accelerator.prepare(text_encoder2)
+        optimizer, train_dataloader, lr_scheduler = accelerator.prepare(optimizer, train_dataloader, lr_scheduler)
 
     # TextEncoderの出力をキャッシュするときにはCPUへ移動する
     if args.cache_text_encoder_outputs:
diff --git a/train_db.py b/train_db.py
index 888cad25e..d5f47a179 100644
--- a/train_db.py
+++ b/train_db.py
@@ -184,7 +184,7 @@ def train(args):
         batch_size=1,
         shuffle=True,
         collate_fn=collator,
-        num_workers=n_workers,
+        num_workers=n_workers if not args.deepspeed else 1, # To avoid RuntimeError: DataLoader worker exited unexpectedly with exit code 1.
         persistent_workers=args.persistent_data_loader_workers,
     )
 
@@ -214,15 +214,36 @@ def train(args):
         text_encoder.to(weight_dtype)
 
     # acceleratorがなんかよろしくやってくれるらしい
-    if train_text_encoder:
-        unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
-            unet, text_encoder, optimizer, train_dataloader, lr_scheduler
-        )
+    if args.deepspeed:
+        # wrapping model
+        class DeepSpeedModel(torch.nn.Module): 
+            def __init__(self, unet, text_encoder, vae) -> None:
+                super().__init__()
+                self.unet = unet
+                self.text_encoders = self.text_encoder = torch.nn.ModuleList(text_encoder)
+                self.vae = vae
+                
+            def get_models(self):
+                return self.unet, self.text_encoders, self.vae
+
+        unet.to(accelerator.device, dtype=weight_dtype)
+        [t_enc.to(accelerator.device, dtype=weight_dtype) for t_enc in text_encoders]
+        ds_model = DeepSpeedModel(unet, text_encoders, vae)
+        ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(ds_model, optimizer, train_dataloader, lr_scheduler)
+        # Now, ds_model is an instance of DeepSpeedEngine. 
+        unet, text_encoders, vae = ds_model.get_models() # for compatiblility
+        vae.to(vae_dtype) # to avoid explicitly half-vae
+        text_encoder = text_encoders
     else:
-        unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(unet, optimizer, train_dataloader, lr_scheduler)
+        if train_text_encoder:
+            unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+                unet, text_encoder, optimizer, train_dataloader, lr_scheduler
+            )
+        else:
+            unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(unet, optimizer, train_dataloader, lr_scheduler)
 
-    if not train_text_encoder:
-        text_encoder.to(accelerator.device, dtype=weight_dtype)  # to avoid 'cpu' vs 'cuda' error
+        if not train_text_encoder:
+            text_encoder.to(accelerator.device, dtype=weight_dtype)  # to avoid 'cpu' vs 'cuda' error
 
     # 実験的機能：勾配も含めたfp16学習を行う　PyTorchにパッチを当ててfp16でのgrad scaleを有効にする
     if args.full_fp16:
diff --git a/train_network.py b/train_network.py
index 8d102ae8f..05dbe2de7 100644
--- a/train_network.py
+++ b/train_network.py
@@ -353,18 +353,26 @@ def train(self, args):
             batch_size=1,
             shuffle=True,
             collate_fn=collator,
-            num_workers=n_workers,
+            num_workers=n_workers if not args.deepspeed else 1, # To avoid RuntimeError: DataLoader worker exited unexpectedly with exit code 1.
             persistent_workers=args.persistent_data_loader_workers,
         )
 
         # 学習ステップ数を計算する
         if args.max_train_epochs is not None:
-            args.max_train_steps = args.max_train_epochs * math.ceil(
-                len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
-            )
-            accelerator.print(
-                f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
-            )
+            if args.deepspeed:
+                args.max_train_steps = args.max_train_epochs * math.ceil(
+                    len(train_dataloader) / args.gradient_accumulation_steps
+                )
+                accelerator.print(
+                    f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
+                )
+            else:
+                args.max_train_steps = args.max_train_epochs * math.ceil(
+                    len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
+                )
+                accelerator.print(
+                    f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
+                )
 
         # データセット側にも学習ステップを送信
         train_dataset_group.set_max_train_steps(args.max_train_steps)
@@ -409,20 +417,42 @@ def train(self, args):
                 t_enc.text_model.embeddings.to(dtype=(weight_dtype if te_weight_dtype != weight_dtype else te_weight_dtype))
 
         # acceleratorがなんかよろしくやってくれるらしい / accelerator will do something good
-        if train_unet:
-            unet = accelerator.prepare(unet)
+        if args.deepspeed:
+            # wrapping model
+            class DeepSpeedModel(torch.nn.Module): 
+                def __init__(self, unet, text_encoder, vae, network) -> None:
+                    super().__init__()
+                    self.unet = unet
+                    self.text_encoders = self.text_encoder = torch.nn.ModuleList(text_encoder)
+                    self.vae = vae
+                    self.network = network
+                    
+                def get_models(self):
+                    return self.unet, self.text_encoders, self.vae, self.network
+            
+            unet.to(accelerator.device, dtype=unet_weight_dtype)
+            [t_enc.to(accelerator.device, dtype=te_weight_dtype) for t_enc in text_encoders]
+            ds_model = DeepSpeedModel(unet, text_encoders, vae, network)
+            ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(ds_model, optimizer, train_dataloader, lr_scheduler)
+            # Now, ds_model is an instance of DeepSpeedEngine. 
+            unet, text_encoders, vae, network = ds_model.get_models() # for compatiblility
+            vae.to(vae_dtype) # to avoid explicitly half-vae
+            text_encoder = text_encoders
         else:
-            unet.to(accelerator.device, dtype=unet_weight_dtype)  # move to device because unet is not prepared by accelerator
-        if train_text_encoder:
-            if len(text_encoders) > 1:
-                text_encoder = text_encoders = [accelerator.prepare(t_enc) for t_enc in text_encoders]
+            if train_unet:
+                unet = accelerator.prepare(unet)
             else:
-                text_encoder = accelerator.prepare(text_encoder)
-                text_encoders = [text_encoder]
-        else:
-            pass  # if text_encoder is not trained, no need to prepare. and device and dtype are already set
+                unet.to(accelerator.device, dtype=unet_weight_dtype)  # move to device because unet is not prepared by accelerator
+            if train_text_encoder:
+                if len(text_encoders) > 1:
+                    text_encoder = text_encoders = [accelerator.prepare(t_enc) for t_enc in text_encoders]
+                else:
+                    text_encoder = accelerator.prepare(text_encoder)
+                    text_encoders = [text_encoder]
+            else:
+                pass  # if text_encoder is not trained, no need to prepare. and device and dtype are already set
 
-        network, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(network, optimizer, train_dataloader, lr_scheduler)
+            network, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(network, optimizer, train_dataloader, lr_scheduler)
 
         if args.gradient_checkpointing:
             # according to TI example in Diffusers, train is required

From 64873c1b4317afad99a1d397454ba0c64c6cb0b1 Mon Sep 17 00:00:00 2001
From: BootsofLagrangian <hard2251@yonsei.ac.kr>
Date: Mon, 5 Feb 2024 17:11:50 +0900
Subject: [PATCH 02/69] fix offload_optimizer_device typo

---
 library/train_util.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/library/train_util.py b/library/train_util.py
index 2d85c9776..933a34c48 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -3134,7 +3134,7 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
         help="Possible options are 0,1,2,3."
     )
     parser.add_argument(
-        "--offload_optimizer", 
+        "--offload_optimizer_device", 
         type=str, default=None,
         choices=[None, "cpu", "nvme"],
         help="Possible options are none|cpu|nvme. Only applicable with ZeRO Stages 2 and 3."
@@ -3959,7 +3959,7 @@ def prepare_accelerator(args: argparse.Namespace):
         deepspeed_plugin = DeepSpeedPlugin(
             zero_stage=args.zero_stage,
             gradient_accumulation_steps=args.gradient_accumulation_steps, gradient_clipping=args.max_grad_norm,
-            offload_optimizer=args.offload_optimizer, offload_optimizer_nvme_path=args.offload_optimizer_nvme_path,
+            offload_optimizer_device=args.offload_optimizer_device, offload_optimizer_nvme_path=args.offload_optimizer_nvme_path,
             offload_param_device=args.offload_param_device, offload_param_nvme_path=args.offload_param_nvme_path,
             zero3_init_flag=args.zero3_init_flag, zero3_save_16bit_model=args.zero3_save_16bit_model,
         )

From 2824312d5eb6df118d7585cde7e84d4cdae6f6c6 Mon Sep 17 00:00:00 2001
From: BootsofLagrangian <hard2251@yonsei.ac.kr>
Date: Mon, 5 Feb 2024 20:13:28 +0900
Subject: [PATCH 03/69] fix vae type error during training sdxl

---
 library/sdxl_train_util.py |  1 -
 library/train_util.py      |  5 -----
 sdxl_train.py              | 25 +++++++++++--------------
 3 files changed, 11 insertions(+), 20 deletions(-)

diff --git a/library/sdxl_train_util.py b/library/sdxl_train_util.py
index 5ad748d15..ff7fef176 100644
--- a/library/sdxl_train_util.py
+++ b/library/sdxl_train_util.py
@@ -17,7 +17,6 @@
 
 
 def load_target_model(args, accelerator, model_version: str, weight_dtype):
-    # load models for each process
     model_dtype = match_mixed_precision(args, weight_dtype)  # prepare fp16/bf16
     for pi in range(accelerator.state.num_processes):
         if pi == accelerator.state.local_process_index:
diff --git a/library/train_util.py b/library/train_util.py
index 933a34c48..a20edbe15 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -4042,28 +4042,23 @@ def _load_target_model(args: argparse.Namespace, weight_dtype, device="cpu", une
 
 
 def load_target_model(args, weight_dtype, accelerator, unet_use_linear_projection_in_v2=False):
-    # load models for each process
     for pi in range(accelerator.state.num_processes):
         if pi == accelerator.state.local_process_index:
             print(f"loading model for process {accelerator.state.local_process_index}/{accelerator.state.num_processes}")
-
             text_encoder, vae, unet, load_stable_diffusion_format = _load_target_model(
                 args,
                 weight_dtype,
                 accelerator.device if args.lowram else "cpu",
                 unet_use_linear_projection_in_v2=unet_use_linear_projection_in_v2,
             )
-
             # work on low-ram device
             if args.lowram:
                 text_encoder.to(accelerator.device)
                 unet.to(accelerator.device)
                 vae.to(accelerator.device)
-
             gc.collect()
             torch.cuda.empty_cache()
         accelerator.wait_for_everyone()
-
     return text_encoder, vae, unet, load_stable_diffusion_format
 
 
diff --git a/sdxl_train.py b/sdxl_train.py
index 6ce6c201e..e8680828b 100644
--- a/sdxl_train.py
+++ b/sdxl_train.py
@@ -392,23 +392,20 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
     if args.deepspeed:
         # Wrapping model for DeepSpeed
         class DeepSpeedModel(torch.nn.Module): 
-            def __init__(self, unet, text_encoder, vae) -> None:
+            def __init__(self, unet, text_encoder) -> None:
                 super().__init__()
                 self.unet = unet
                 self.text_encoders = self.text_encoder = torch.nn.ModuleList(text_encoder)
-                self.vae = vae
                 
             def get_models(self):
-                return self.unet, self.text_encoders, self.vae
+                return self.unet, self.text_encoders
         text_encoders = [text_encoder1, text_encoder2]
-        unet.to(accelerator.device, dtype=weight_dtype)
-        [t_enc.to(accelerator.device, dtype=weight_dtype) for t_enc in text_encoders]
-        ds_model = DeepSpeedModel(unet, text_encoders, vae)
+        ds_model = DeepSpeedModel(unet, text_encoders)
         ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(ds_model, optimizer, train_dataloader, lr_scheduler)
         # Now, ds_model is an instance of DeepSpeedEngine. 
-        unet, text_encoders, vae = ds_model.get_models() # for compatiblility
-        vae.to(vae_dtype) # to avoid explicitly half-vae
-        text_encoder1, text_encoder2 = text_encoders[0], text_encoders[1]
+        unet, text_encoders = ds_model.get_models() # for compatiblility
+        text_encoder1, text_encoder2 = text_encoder = text_encoders
+        training_models = [unet, text_encoder1, text_encoder2]
     else: # acceleratorがなんかよろしくやってくれるらしい
         if train_unet:
             unet = accelerator.prepare(unet)
@@ -493,10 +490,10 @@ def get_models(self):
         for step, batch in enumerate(train_dataloader):
             current_step.value = global_step
             with accelerator.accumulate(*training_models):
-                if "latents" in batch and batch["latents"] is not None:
-                    latents = batch["latents"].to(accelerator.device).to(dtype=weight_dtype)
-                else:
-                    with torch.no_grad():
+                with torch.no_grad(): # why this block differ within train_network.py?
+                    if "latents" in batch and batch["latents"] is not None:
+                        latents = batch["latents"].to(accelerator.device).to(dtype=weight_dtype)
+                    else:
                         # latentに変換
                         latents = vae.encode(batch["images"].to(vae_dtype)).latent_dist.sample().to(weight_dtype)
 
@@ -504,7 +501,7 @@ def get_models(self):
                         if torch.any(torch.isnan(latents)):
                             accelerator.print("NaN found in latents, replacing with zeros")
                             latents = torch.nan_to_num(latents, 0, out=latents)
-                latents = latents * sdxl_model_util.VAE_SCALE_FACTOR
+                    latents = latents * sdxl_model_util.VAE_SCALE_FACTOR
 
                 if "text_encoder_outputs1_list" not in batch or batch["text_encoder_outputs1_list"] is None:
                     input_ids1 = batch["input_ids"]

From 4295f91dcd75a7405aa70d5c5d2c826a618a4bcc Mon Sep 17 00:00:00 2001
From: BootsofLagrangian <hard2251@yonsei.ac.kr>
Date: Mon, 5 Feb 2024 20:19:56 +0900
Subject: [PATCH 04/69] fix all trainer about vae

---
 fine_tune.py     | 29 ++++++++++++++++-------------
 train_db.py      | 29 ++++++++++++++++-------------
 train_network.py | 15 +++++----------
 3 files changed, 37 insertions(+), 36 deletions(-)

diff --git a/fine_tune.py b/fine_tune.py
index 78dfd1696..f901ee641 100644
--- a/fine_tune.py
+++ b/fine_tune.py
@@ -221,10 +221,18 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
 
     # 学習ステップ数を計算する
     if args.max_train_epochs is not None:
-        args.max_train_steps = args.max_train_epochs * math.ceil(
-            len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
-        )
-        accelerator.print(f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}")
+        if args.deepspeed:
+            args.max_train_steps = args.max_train_epochs * math.ceil(
+                len(train_dataloader) / args.gradient_accumulation_steps
+            )
+            accelerator.print(
+                f"[DeepSpeed] override steps not dividing by {accelerator.num_processes}. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
+            )
+        else:
+            args.max_train_steps = args.max_train_epochs * math.ceil(
+                len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
+            )
+            accelerator.print(f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}")
 
     # データセット側にも学習ステップを送信
     train_dataset_group.set_max_train_steps(args.max_train_steps)
@@ -244,21 +252,16 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
     if args.deepspeed:
         # wrapping model
         class DeepSpeedModel(torch.nn.Module): 
-            def __init__(self, unet, text_encoder, vae) -> None:
+            def __init__(self, unet, text_encoder) -> None:
                 super().__init__()
                 self.unet = unet
                 self.text_encoders = self.text_encoder = torch.nn.ModuleList(text_encoder)
-                self.vae = vae
             def get_models(self):
-                return self.unet, self.text_encoders, self.vae
-            
-        unet.to(accelerator.device, dtype=weight_dtype)
-        [t_enc.to(accelerator.device, dtype=weight_dtype) for t_enc in text_encoders]
-        ds_model = DeepSpeedModel(unet, text_encoders, vae)
+                return self.unet, self.text_encoders
+        ds_model = DeepSpeedModel(unet, text_encoders)
         ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(ds_model, optimizer, train_dataloader, lr_scheduler)
         # Now, ds_model is an instance of DeepSpeedEngine. 
-        unet, text_encoders, vae = ds_model.get_models() # for compatiblility
-        vae.to(vae_dtype)
+        unet, text_encoders = ds_model.get_models() # for compatiblility
         text_encoder = text_encoders
             
     else: # acceleratorがなんかよろしくやってくれるらしい
diff --git a/train_db.py b/train_db.py
index d5f47a179..fa7f6a8dc 100644
--- a/train_db.py
+++ b/train_db.py
@@ -190,10 +190,18 @@ def train(args):
 
     # 学習ステップ数を計算する
     if args.max_train_epochs is not None:
-        args.max_train_steps = args.max_train_epochs * math.ceil(
-            len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
-        )
-        accelerator.print(f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}")
+        if args.deepspeed:
+            args.max_train_steps = args.max_train_epochs * math.ceil(
+                len(train_dataloader) / args.gradient_accumulation_steps
+            )
+            accelerator.print(
+                f"[DeepSpeed] override steps not dividing by {accelerator.num_processes}. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
+            )
+        else:
+            args.max_train_steps = args.max_train_epochs * math.ceil(
+                len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
+            )
+            accelerator.print(f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}")
 
     # データセット側にも学習ステップを送信
     train_dataset_group.set_max_train_steps(args.max_train_steps)
@@ -217,22 +225,17 @@ def train(args):
     if args.deepspeed:
         # wrapping model
         class DeepSpeedModel(torch.nn.Module): 
-            def __init__(self, unet, text_encoder, vae) -> None:
+            def __init__(self, unet, text_encoder) -> None:
                 super().__init__()
                 self.unet = unet
                 self.text_encoders = self.text_encoder = torch.nn.ModuleList(text_encoder)
-                self.vae = vae
                 
             def get_models(self):
-                return self.unet, self.text_encoders, self.vae
-
-        unet.to(accelerator.device, dtype=weight_dtype)
-        [t_enc.to(accelerator.device, dtype=weight_dtype) for t_enc in text_encoders]
-        ds_model = DeepSpeedModel(unet, text_encoders, vae)
+                return self.unet, self.text_encoders
+        ds_model = DeepSpeedModel(unet, text_encoders)
         ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(ds_model, optimizer, train_dataloader, lr_scheduler)
         # Now, ds_model is an instance of DeepSpeedEngine. 
-        unet, text_encoders, vae = ds_model.get_models() # for compatiblility
-        vae.to(vae_dtype) # to avoid explicitly half-vae
+        unet, text_encoders = ds_model.get_models() # for compatiblility
         text_encoder = text_encoders
     else:
         if train_text_encoder:
diff --git a/train_network.py b/train_network.py
index 05dbe2de7..bbda427aa 100644
--- a/train_network.py
+++ b/train_network.py
@@ -364,7 +364,7 @@ def train(self, args):
                     len(train_dataloader) / args.gradient_accumulation_steps
                 )
                 accelerator.print(
-                    f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
+                    f"[DeepSpeed] override steps not dividing by {accelerator.num_processes}. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
                 )
             else:
                 args.max_train_steps = args.max_train_epochs * math.ceil(
@@ -420,23 +420,18 @@ def train(self, args):
         if args.deepspeed:
             # wrapping model
             class DeepSpeedModel(torch.nn.Module): 
-                def __init__(self, unet, text_encoder, vae, network) -> None:
+                def __init__(self, unet, text_encoder, network) -> None:
                     super().__init__()
                     self.unet = unet
                     self.text_encoders = self.text_encoder = torch.nn.ModuleList(text_encoder)
-                    self.vae = vae
                     self.network = network
                     
                 def get_models(self):
-                    return self.unet, self.text_encoders, self.vae, self.network
-            
-            unet.to(accelerator.device, dtype=unet_weight_dtype)
-            [t_enc.to(accelerator.device, dtype=te_weight_dtype) for t_enc in text_encoders]
-            ds_model = DeepSpeedModel(unet, text_encoders, vae, network)
+                    return self.unet, self.text_encoders, self.network
+            ds_model = DeepSpeedModel(unet, text_encoders, network)
             ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(ds_model, optimizer, train_dataloader, lr_scheduler)
             # Now, ds_model is an instance of DeepSpeedEngine. 
-            unet, text_encoders, vae, network = ds_model.get_models() # for compatiblility
-            vae.to(vae_dtype) # to avoid explicitly half-vae
+            unet, text_encoders, network = ds_model.get_models() # for compatiblility
             text_encoder = text_encoders
         else:
             if train_unet:

From 3970bf40804d9c66e76e0af5e1d0477f19bfa79a Mon Sep 17 00:00:00 2001
From: BootsofLagrangian <hard2251@yonsei.ac.kr>
Date: Mon, 5 Feb 2024 22:40:43 +0900
Subject: [PATCH 05/69] maybe fix branch to run offloading

---
 library/train_util.py | 2 ++
 sdxl_train.py         | 6 ++++++
 2 files changed, 8 insertions(+)

diff --git a/library/train_util.py b/library/train_util.py
index a20edbe15..676652e90 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -3964,6 +3964,8 @@ def prepare_accelerator(args: argparse.Namespace):
             zero3_init_flag=args.zero3_init_flag, zero3_save_16bit_model=args.zero3_save_16bit_model,
         )
         deepspeed_plugin.deepspeed_config['train_micro_batch_size_per_gpu'] = args.train_batch_size
+        deepspeed_plugin.deepspeed_config['train_batch_size'] = \
+            args.train_batch_size * args.gradient_accumulation_steps * int(os.environ['WORLD_SIZE'])
 
     accelerator = Accelerator(
         gradient_accumulation_steps=args.gradient_accumulation_steps,
diff --git a/sdxl_train.py b/sdxl_train.py
index e8680828b..ef3ead380 100644
--- a/sdxl_train.py
+++ b/sdxl_train.py
@@ -391,6 +391,12 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
 
     if args.deepspeed:
         # Wrapping model for DeepSpeed
+        import deepspeed
+        if args.offload_optimizer_device is not None:
+            accelerator.print('[DeepSpeed] start to manually build cpu_adam.')
+            deepspeed.ops.op_builder.CPUAdamBuilder().load()
+            accelerator.print('[DeepSpeed] building cpu_adam done.')
+
         class DeepSpeedModel(torch.nn.Module): 
             def __init__(self, unet, text_encoder) -> None:
                 super().__init__()

From 7d2a9268b9d8d3c9b78068aaa2f9d43eb8b6101b Mon Sep 17 00:00:00 2001
From: BootsofLagrangian <hard2251@yonsei.ac.kr>
Date: Mon, 5 Feb 2024 22:42:06 +0900
Subject: [PATCH 06/69] apply offloading method runable for all trainer

---
 fine_tune.py     | 5 +++++
 train_db.py      | 5 +++++
 train_network.py | 5 +++++
 3 files changed, 15 insertions(+)

diff --git a/fine_tune.py b/fine_tune.py
index f901ee641..85febeaaa 100644
--- a/fine_tune.py
+++ b/fine_tune.py
@@ -251,6 +251,11 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
 
     if args.deepspeed:
         # wrapping model
+        import deepspeed
+        if args.offload_optimizer_device is not None:
+            accelerator.print('[DeepSpeed] start to manually build cpu_adam.')
+            deepspeed.ops.op_builder.CPUAdamBuilder().load()
+            accelerator.print('[DeepSpeed] building cpu_adam done.')
         class DeepSpeedModel(torch.nn.Module): 
             def __init__(self, unet, text_encoder) -> None:
                 super().__init__()
diff --git a/train_db.py b/train_db.py
index fa7f6a8dc..e26618867 100644
--- a/train_db.py
+++ b/train_db.py
@@ -224,6 +224,11 @@ def train(args):
     # acceleratorがなんかよろしくやってくれるらしい
     if args.deepspeed:
         # wrapping model
+        import deepspeed
+        if args.offload_optimizer_device is not None:
+            accelerator.print('[DeepSpeed] start to manually build cpu_adam.')
+            deepspeed.ops.op_builder.CPUAdamBuilder().load()
+            accelerator.print('[DeepSpeed] building cpu_adam done.')
         class DeepSpeedModel(torch.nn.Module): 
             def __init__(self, unet, text_encoder) -> None:
                 super().__init__()
diff --git a/train_network.py b/train_network.py
index bbda427aa..050a65111 100644
--- a/train_network.py
+++ b/train_network.py
@@ -419,6 +419,11 @@ def train(self, args):
         # acceleratorがなんかよろしくやってくれるらしい / accelerator will do something good
         if args.deepspeed:
             # wrapping model
+            import deepspeed
+            if args.offload_optimizer_device is not None:
+                accelerator.print('[DeepSpeed] start to manually build cpu_adam.')
+                deepspeed.ops.op_builder.CPUAdamBuilder().load()
+                accelerator.print('[DeepSpeed] building cpu_adam done.')
             class DeepSpeedModel(torch.nn.Module): 
                 def __init__(self, unet, text_encoder, network) -> None:
                     super().__init__()

From 62556619bdc876c450bfb1445b16683cf3a98699 Mon Sep 17 00:00:00 2001
From: BootsofLagrangian <hard2251@yonsei.ac.kr>
Date: Wed, 7 Feb 2024 16:42:05 +0900
Subject: [PATCH 07/69] fix full_fp16 compatible and train_step

---
 fine_tune.py              | 16 ++-----
 library/train_util.py     |  7 +++
 sdxl_train.py             |  3 +-
 test_pip_requirements.txt | 96 +++++++++++++++++++++++++++++++++++++++
 train_db.py               | 16 ++-----
 train_network.py          | 23 ++++------
 6 files changed, 121 insertions(+), 40 deletions(-)
 create mode 100644 test_pip_requirements.txt

diff --git a/fine_tune.py b/fine_tune.py
index 85febeaaa..eb652742c 100644
--- a/fine_tune.py
+++ b/fine_tune.py
@@ -221,18 +221,10 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
 
     # 学習ステップ数を計算する
     if args.max_train_epochs is not None:
-        if args.deepspeed:
-            args.max_train_steps = args.max_train_epochs * math.ceil(
-                len(train_dataloader) / args.gradient_accumulation_steps
-            )
-            accelerator.print(
-                f"[DeepSpeed] override steps not dividing by {accelerator.num_processes}. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
-            )
-        else:
-            args.max_train_steps = args.max_train_epochs * math.ceil(
-                len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
-            )
-            accelerator.print(f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}")
+        args.max_train_steps = args.max_train_epochs * math.ceil(
+            len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
+        )
+        accelerator.print(f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}")
 
     # データセット側にも学習ステップを送信
     train_dataset_group.set_max_train_steps(args.max_train_steps)
diff --git a/library/train_util.py b/library/train_util.py
index 676652e90..ea6265109 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -3166,6 +3166,11 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
         action="store_true",
         help="Flag to indicate whether to save 16-bit model. Only applicable with ZeRO Stage-3."
     )
+    parser.add_argument(
+        "--fp16_master_weights_and_gradients",
+        action="store_true",
+        help="fp16_master_and_gradients requires optimizer to support keeping fp16 master and gradients while keeping the optimizer states in fp32."
+    )
 
 def verify_training_args(args: argparse.Namespace):
     if args.v_parameterization and not args.v2:
@@ -3966,6 +3971,8 @@ def prepare_accelerator(args: argparse.Namespace):
         deepspeed_plugin.deepspeed_config['train_micro_batch_size_per_gpu'] = args.train_batch_size
         deepspeed_plugin.deepspeed_config['train_batch_size'] = \
             args.train_batch_size * args.gradient_accumulation_steps * int(os.environ['WORLD_SIZE'])
+        if args.full_fp16 or args.fp16_master_weights_and_gradients:
+            deepspeed_plugin.deepspeed_config['fp16_master_weights_and_gradients'] = True
 
     accelerator = Accelerator(
         gradient_accumulation_steps=args.gradient_accumulation_steps,
diff --git a/sdxl_train.py b/sdxl_train.py
index ef3ead380..54902b873 100644
--- a/sdxl_train.py
+++ b/sdxl_train.py
@@ -437,7 +437,8 @@ def get_models(self):
         text_encoder2.to(accelerator.device)
 
     # 実験的機能：勾配も含めたfp16学習を行う　PyTorchにパッチを当ててfp16でのgrad scaleを有効にする
-    if args.full_fp16:
+    if args.full_fp16 and not args.deepspeed:
+        # During deepseed training, accelerate not handles fp16/bf16|mixed precision directly via scaler. Let deepspeed engine do.
         train_util.patch_accelerator_for_fp16_training(accelerator)
 
     # resumeする
diff --git a/test_pip_requirements.txt b/test_pip_requirements.txt
new file mode 100644
index 000000000..6abec3516
--- /dev/null
+++ b/test_pip_requirements.txt
@@ -0,0 +1,96 @@
+absl-py==2.1.0
+accelerate==0.25.0
+aiohttp==3.9.3
+aiosignal==1.3.1
+altair==4.2.2
+annotated-types @ file:///home/conda/feedstock_root/build_artifacts/annotated-types_1696634205638/work
+async-timeout==4.0.3
+attrs==23.2.0
+bitsandbytes==0.42.0
+Brotli @ file:///home/conda/feedstock_root/build_artifacts/brotli-split_1695989787169/work
+cachetools==5.3.2
+certifi==2022.12.7
+charset-normalizer==2.1.1
+cmake==3.25.0
+deepspeed==0.13.1
+diffusers==0.25.0
+easygui==0.98.3
+einops==0.6.1
+entrypoints==0.4
+filelock==3.9.0
+frozenlist==1.4.1
+fsspec==2024.2.0
+ftfy==6.1.1
+gmpy2 @ file:///home/conda/feedstock_root/build_artifacts/gmpy2_1666808654411/work
+google-auth==2.27.0
+google-auth-oauthlib==0.4.6
+grpcio==1.60.1
+hjson==3.1.0
+huggingface-hub==0.20.1
+idna==3.4
+importlib-metadata==7.0.1
+Jinja2==3.1.2
+jsonschema==4.21.1
+jsonschema-specifications==2023.12.1
+-e git+https://github.com/kohya-ss/sd-scripts@cd19df49cd512e13ac90db115c424d19c0e8868a#egg=library
+lightning-utilities==0.10.1
+lit==15.0.7
+Markdown==3.5.2
+MarkupSafe==2.1.3
+mpmath==1.3.0
+multidict==6.0.5
+networkx==3.2.1
+ninja==1.11.1.1
+numpy==1.26.3
+oauthlib==3.2.2
+open-clip-torch==2.20.0
+opencv-python==4.7.0.68
+packaging==23.2
+pandas==2.2.0
+pillow==10.2.0
+protobuf==3.19.6
+psutil==5.9.8
+py-cpuinfo @ file:///home/conda/feedstock_root/build_artifacts/py-cpuinfo_1666774466606/work
+pyasn1==0.5.1
+pyasn1-modules==0.3.0
+pydantic @ file:///home/conda/feedstock_root/build_artifacts/pydantic_1706543943340/work
+pydantic_core @ file:///home/conda/feedstock_root/build_artifacts/pydantic-core_1705674688239/work
+pynvml==11.5.0
+PySocks @ file:///home/conda/feedstock_root/build_artifacts/pysocks_1661604839144/work
+python-dateutil==2.8.2
+pytorch-lightning==1.9.0
+pytz==2024.1
+PyYAML @ file:///home/conda/feedstock_root/build_artifacts/pyyaml_1695373428874/work
+referencing==0.33.0
+regex==2023.12.25
+requests==2.28.1
+requests-oauthlib==1.3.1
+rpds-py==0.17.1
+rsa==4.9
+safetensors==0.4.2
+scipy==1.12.0
+sentencepiece==0.1.99
+six==1.16.0
+sympy==1.12
+tensorboard==2.10.1
+tensorboard-data-server==0.6.1
+tensorboard-plugin-wit==1.8.1
+timm==0.9.12
+tokenizers==0.15.1
+toml==0.10.2
+toolz==0.12.1
+torch==2.0.1+cu118
+torchaudio==2.2.0
+torchmetrics==1.3.0.post0
+torchvision==0.15.2+cu118
+tqdm==4.66.1
+transformers==4.36.2
+triton==2.0.0
+typing_extensions==4.8.0
+tzdata==2023.4
+urllib3==1.26.13
+voluptuous==0.13.1
+wcwidth==0.2.13
+Werkzeug==3.0.1
+yarl==1.9.4
+zipp==3.17.0
diff --git a/train_db.py b/train_db.py
index e26618867..58536555e 100644
--- a/train_db.py
+++ b/train_db.py
@@ -190,18 +190,10 @@ def train(args):
 
     # 学習ステップ数を計算する
     if args.max_train_epochs is not None:
-        if args.deepspeed:
-            args.max_train_steps = args.max_train_epochs * math.ceil(
-                len(train_dataloader) / args.gradient_accumulation_steps
-            )
-            accelerator.print(
-                f"[DeepSpeed] override steps not dividing by {accelerator.num_processes}. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
-            )
-        else:
-            args.max_train_steps = args.max_train_epochs * math.ceil(
-                len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
-            )
-            accelerator.print(f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}")
+        args.max_train_steps = args.max_train_epochs * math.ceil(
+            len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
+        )
+        accelerator.print(f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}")
 
     # データセット側にも学習ステップを送信
     train_dataset_group.set_max_train_steps(args.max_train_steps)
diff --git a/train_network.py b/train_network.py
index 050a65111..cc445d39a 100644
--- a/train_network.py
+++ b/train_network.py
@@ -359,20 +359,12 @@ def train(self, args):
 
         # 学習ステップ数を計算する
         if args.max_train_epochs is not None:
-            if args.deepspeed:
-                args.max_train_steps = args.max_train_epochs * math.ceil(
-                    len(train_dataloader) / args.gradient_accumulation_steps
-                )
-                accelerator.print(
-                    f"[DeepSpeed] override steps not dividing by {accelerator.num_processes}. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
-                )
-            else:
-                args.max_train_steps = args.max_train_epochs * math.ceil(
-                    len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
-                )
-                accelerator.print(
-                    f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
-                )
+            args.max_train_steps = args.max_train_epochs * math.ceil(
+                len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
+            )
+            accelerator.print(
+                f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
+            )
 
         # データセット側にも学習ステップを送信
         train_dataset_group.set_max_train_steps(args.max_train_steps)
@@ -479,7 +471,8 @@ def get_models(self):
             vae.to(accelerator.device, dtype=vae_dtype)
 
         # 実験的機能：勾配も含めたfp16学習を行う　PyTorchにパッチを当ててfp16でのgrad scaleを有効にする
-        if args.full_fp16:
+        if args.full_fp16 and not args.deepspeed:
+            # During deepseed training, accelerate not handles fp16/bf16|mixed precision directly via scaler. Let deepspeed engine do.
             train_util.patch_accelerator_for_fp16_training(accelerator)
 
         # resumeする

From 2445a5b74e4c5bb0af24e0b3162c1eaef218b56b Mon Sep 17 00:00:00 2001
From: BootsofLagrangian <hard2251@yonsei.ac.kr>
Date: Wed, 7 Feb 2024 16:48:18 +0900
Subject: [PATCH 08/69] remove test requirements

---
 test_pip_requirements.txt | 96 ---------------------------------------
 1 file changed, 96 deletions(-)
 delete mode 100644 test_pip_requirements.txt

diff --git a/test_pip_requirements.txt b/test_pip_requirements.txt
deleted file mode 100644
index 6abec3516..000000000
--- a/test_pip_requirements.txt
+++ /dev/null
@@ -1,96 +0,0 @@
-absl-py==2.1.0
-accelerate==0.25.0
-aiohttp==3.9.3
-aiosignal==1.3.1
-altair==4.2.2
-annotated-types @ file:///home/conda/feedstock_root/build_artifacts/annotated-types_1696634205638/work
-async-timeout==4.0.3
-attrs==23.2.0
-bitsandbytes==0.42.0
-Brotli @ file:///home/conda/feedstock_root/build_artifacts/brotli-split_1695989787169/work
-cachetools==5.3.2
-certifi==2022.12.7
-charset-normalizer==2.1.1
-cmake==3.25.0
-deepspeed==0.13.1
-diffusers==0.25.0
-easygui==0.98.3
-einops==0.6.1
-entrypoints==0.4
-filelock==3.9.0
-frozenlist==1.4.1
-fsspec==2024.2.0
-ftfy==6.1.1
-gmpy2 @ file:///home/conda/feedstock_root/build_artifacts/gmpy2_1666808654411/work
-google-auth==2.27.0
-google-auth-oauthlib==0.4.6
-grpcio==1.60.1
-hjson==3.1.0
-huggingface-hub==0.20.1
-idna==3.4
-importlib-metadata==7.0.1
-Jinja2==3.1.2
-jsonschema==4.21.1
-jsonschema-specifications==2023.12.1
--e git+https://github.com/kohya-ss/sd-scripts@cd19df49cd512e13ac90db115c424d19c0e8868a#egg=library
-lightning-utilities==0.10.1
-lit==15.0.7
-Markdown==3.5.2
-MarkupSafe==2.1.3
-mpmath==1.3.0
-multidict==6.0.5
-networkx==3.2.1
-ninja==1.11.1.1
-numpy==1.26.3
-oauthlib==3.2.2
-open-clip-torch==2.20.0
-opencv-python==4.7.0.68
-packaging==23.2
-pandas==2.2.0
-pillow==10.2.0
-protobuf==3.19.6
-psutil==5.9.8
-py-cpuinfo @ file:///home/conda/feedstock_root/build_artifacts/py-cpuinfo_1666774466606/work
-pyasn1==0.5.1
-pyasn1-modules==0.3.0
-pydantic @ file:///home/conda/feedstock_root/build_artifacts/pydantic_1706543943340/work
-pydantic_core @ file:///home/conda/feedstock_root/build_artifacts/pydantic-core_1705674688239/work
-pynvml==11.5.0
-PySocks @ file:///home/conda/feedstock_root/build_artifacts/pysocks_1661604839144/work
-python-dateutil==2.8.2
-pytorch-lightning==1.9.0
-pytz==2024.1
-PyYAML @ file:///home/conda/feedstock_root/build_artifacts/pyyaml_1695373428874/work
-referencing==0.33.0
-regex==2023.12.25
-requests==2.28.1
-requests-oauthlib==1.3.1
-rpds-py==0.17.1
-rsa==4.9
-safetensors==0.4.2
-scipy==1.12.0
-sentencepiece==0.1.99
-six==1.16.0
-sympy==1.12
-tensorboard==2.10.1
-tensorboard-data-server==0.6.1
-tensorboard-plugin-wit==1.8.1
-timm==0.9.12
-tokenizers==0.15.1
-toml==0.10.2
-toolz==0.12.1
-torch==2.0.1+cu118
-torchaudio==2.2.0
-torchmetrics==1.3.0.post0
-torchvision==0.15.2+cu118
-tqdm==4.66.1
-transformers==4.36.2
-triton==2.0.0
-typing_extensions==4.8.0
-tzdata==2023.4
-urllib3==1.26.13
-voluptuous==0.13.1
-wcwidth==0.2.13
-Werkzeug==3.0.1
-yarl==1.9.4
-zipp==3.17.0

From a98fecaeb1e818c778c90fe441a71a8bd34615ff Mon Sep 17 00:00:00 2001
From: BootsofLagrangian <hard2251@yonsei.ac.kr>
Date: Wed, 7 Feb 2024 17:19:46 +0900
Subject: [PATCH 09/69] forgot setting mixed_precision for deepspeed. sorry

---
 library/train_util.py | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/library/train_util.py b/library/train_util.py
index ea6265109..dbe5a61ce 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -3971,6 +3971,9 @@ def prepare_accelerator(args: argparse.Namespace):
         deepspeed_plugin.deepspeed_config['train_micro_batch_size_per_gpu'] = args.train_batch_size
         deepspeed_plugin.deepspeed_config['train_batch_size'] = \
             args.train_batch_size * args.gradient_accumulation_steps * int(os.environ['WORLD_SIZE'])
+        deepspeed_plugin.set_mixed_precision(args.mixed_precision)
+        if args.mixed_precision.lower() == "fp16":
+            deepspeed_plugin.deepspeed_config['fp16']['initial_scale_power'] = 0
         if args.full_fp16 or args.fp16_master_weights_and_gradients:
             deepspeed_plugin.deepspeed_config['fp16_master_weights_and_gradients'] = True
 

From 03f0816f86b2d4d8915d81146242fb6f7f99c5ff Mon Sep 17 00:00:00 2001
From: BootsofLagrangian <hard2251@yonsei.ac.kr>
Date: Fri, 9 Feb 2024 17:47:49 +0900
Subject: [PATCH 10/69] the reason not working grad accum steps found. it was
 becasue of my accelerate settings

---
 fine_tune.py          | 5 +++--
 library/train_util.py | 6 +++++-
 sdxl_train.py         | 5 +++--
 train_db.py           | 5 +++--
 4 files changed, 14 insertions(+), 7 deletions(-)

diff --git a/fine_tune.py b/fine_tune.py
index eb652742c..741e9c857 100644
--- a/fine_tune.py
+++ b/fine_tune.py
@@ -224,8 +224,9 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
         args.max_train_steps = args.max_train_epochs * math.ceil(
             len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
         )
-        accelerator.print(f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}")
-
+        accelerator.print(
+            f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
+        )
     # データセット側にも学習ステップを送信
     train_dataset_group.set_max_train_steps(args.max_train_steps)
 
diff --git a/library/train_util.py b/library/train_util.py
index dbe5a61ce..61c836247 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -3975,7 +3975,11 @@ def prepare_accelerator(args: argparse.Namespace):
         if args.mixed_precision.lower() == "fp16":
             deepspeed_plugin.deepspeed_config['fp16']['initial_scale_power'] = 0
         if args.full_fp16 or args.fp16_master_weights_and_gradients:
-            deepspeed_plugin.deepspeed_config['fp16_master_weights_and_gradients'] = True
+            if args.offload_optimizer_device == "cpu":
+                deepspeed_plugin.deepspeed_config['fp16']['fp16_master_weights_and_grads'] = True
+                print("[DeepSpeed] full fp16 enable.")
+            else:
+                print("full fp16, fp16_master_weights_and_grads currently only supported using ZeRO-Offload with DeepSpeedCPUAdam.")
 
     accelerator = Accelerator(
         gradient_accumulation_steps=args.gradient_accumulation_steps,
diff --git a/sdxl_train.py b/sdxl_train.py
index 54902b873..6ffb1bbaf 100644
--- a/sdxl_train.py
+++ b/sdxl_train.py
@@ -363,8 +363,9 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
         args.max_train_steps = args.max_train_epochs * math.ceil(
             len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
         )
-        accelerator.print(f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}")
-
+        accelerator.print(
+            f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
+        )
     # データセット側にも学習ステップを送信
     train_dataset_group.set_max_train_steps(args.max_train_steps)
 
diff --git a/train_db.py b/train_db.py
index 58536555e..c336a1c1c 100644
--- a/train_db.py
+++ b/train_db.py
@@ -193,8 +193,9 @@ def train(args):
         args.max_train_steps = args.max_train_epochs * math.ceil(
             len(train_dataloader) / accelerator.num_processes / args.gradient_accumulation_steps
         )
-        accelerator.print(f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}")
-
+        accelerator.print(
+            f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
+        )
     # データセット側にも学習ステップを送信
     train_dataset_group.set_max_train_steps(args.max_train_steps)
 

From 4d5186d1cf0b0fbda20513def793ac3f5e9d5ea0 Mon Sep 17 00:00:00 2001
From: BootsofLagrangian <hard2251@yonsei.ac.kr>
Date: Thu, 22 Feb 2024 16:20:53 +0900
Subject: [PATCH 11/69] refactored codes, some function moved into
 train_utils.py

---
 fine_tune.py          | 29 +++++++---------
 library/train_util.py | 78 +++++++++++++++++++++++++++++++------------
 sdxl_train.py         | 43 ++++++++++++------------
 train_db.py           | 31 ++++++++---------
 train_network.py      | 34 +++++++++----------
 5 files changed, 119 insertions(+), 96 deletions(-)

diff --git a/fine_tune.py b/fine_tune.py
index 741e9c857..862607545 100644
--- a/fine_tune.py
+++ b/fine_tune.py
@@ -243,24 +243,19 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
         text_encoder.to(weight_dtype)
 
     if args.deepspeed:
-        # wrapping model
-        import deepspeed
-        if args.offload_optimizer_device is not None:
-            accelerator.print('[DeepSpeed] start to manually build cpu_adam.')
-            deepspeed.ops.op_builder.CPUAdamBuilder().load()
-            accelerator.print('[DeepSpeed] building cpu_adam done.')
-        class DeepSpeedModel(torch.nn.Module): 
-            def __init__(self, unet, text_encoder) -> None:
-                super().__init__()
-                self.unet = unet
-                self.text_encoders = self.text_encoder = torch.nn.ModuleList(text_encoder)
-            def get_models(self):
-                return self.unet, self.text_encoders
-        ds_model = DeepSpeedModel(unet, text_encoders)
+        training_models_dict = {}
+        training_models_dict["unet"] = unet
+        if args.train_text_encoder: training_models_dict["text_encoder"] = text_encoder
+
+        ds_model = train_util.prepare_deepspeed_model(args, **training_models_dict)
         ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(ds_model, optimizer, train_dataloader, lr_scheduler)
-        # Now, ds_model is an instance of DeepSpeedEngine. 
-        unet, text_encoders = ds_model.get_models() # for compatiblility
-        text_encoder = text_encoders
+    
+        training_models = []
+        unet = ds_model.models["unet"]
+        training_models.append(unet)
+        if args.train_text_encoder:
+            text_encoder = ds_model.models["text_encoder"]
+            training_models.append(text_encoder)
             
     else: # acceleratorがなんかよろしくやってくれるらしい
         if args.train_text_encoder:
diff --git a/library/train_util.py b/library/train_util.py
index 61c836247..334aaa21e 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -3959,27 +3959,7 @@ def prepare_accelerator(args: argparse.Namespace):
         else None,
     )
     kwargs_handlers = list(filter(lambda x: x is not None, kwargs_handlers))
-    deepspeed_plugin = None
-    if args.deepspeed:
-        deepspeed_plugin = DeepSpeedPlugin(
-            zero_stage=args.zero_stage,
-            gradient_accumulation_steps=args.gradient_accumulation_steps, gradient_clipping=args.max_grad_norm,
-            offload_optimizer_device=args.offload_optimizer_device, offload_optimizer_nvme_path=args.offload_optimizer_nvme_path,
-            offload_param_device=args.offload_param_device, offload_param_nvme_path=args.offload_param_nvme_path,
-            zero3_init_flag=args.zero3_init_flag, zero3_save_16bit_model=args.zero3_save_16bit_model,
-        )
-        deepspeed_plugin.deepspeed_config['train_micro_batch_size_per_gpu'] = args.train_batch_size
-        deepspeed_plugin.deepspeed_config['train_batch_size'] = \
-            args.train_batch_size * args.gradient_accumulation_steps * int(os.environ['WORLD_SIZE'])
-        deepspeed_plugin.set_mixed_precision(args.mixed_precision)
-        if args.mixed_precision.lower() == "fp16":
-            deepspeed_plugin.deepspeed_config['fp16']['initial_scale_power'] = 0
-        if args.full_fp16 or args.fp16_master_weights_and_gradients:
-            if args.offload_optimizer_device == "cpu":
-                deepspeed_plugin.deepspeed_config['fp16']['fp16_master_weights_and_grads'] = True
-                print("[DeepSpeed] full fp16 enable.")
-            else:
-                print("full fp16, fp16_master_weights_and_grads currently only supported using ZeRO-Offload with DeepSpeedCPUAdam.")
+    deepspeed_plugin = prepare_deepspeed_plugin(args)
 
     accelerator = Accelerator(
         gradient_accumulation_steps=args.gradient_accumulation_steps,
@@ -3992,6 +3972,62 @@ def prepare_accelerator(args: argparse.Namespace):
     )
     return accelerator
 
+def prepare_deepspeed_plugin(args: argparse.Namespace):
+    if args.deepspeed is None: return None
+    try:
+        import deepspeed
+    except ImportError as e:
+        print("deepspeed is not installed. please install deepspeed in your environment with following command. DS_BUILD_OPS=0 pip install deepspeed")
+        exit(1)
+
+    deepspeed_plugin = DeepSpeedPlugin(
+        zero_stage=args.zero_stage,
+        gradient_accumulation_steps=args.gradient_accumulation_steps, gradient_clipping=args.max_grad_norm,
+        offload_optimizer_device=args.offload_optimizer_device, offload_optimizer_nvme_path=args.offload_optimizer_nvme_path,
+        offload_param_device=args.offload_param_device, offload_param_nvme_path=args.offload_param_nvme_path,
+        zero3_init_flag=args.zero3_init_flag, zero3_save_16bit_model=args.zero3_save_16bit_model,
+    )
+    deepspeed_plugin.deepspeed_config['train_micro_batch_size_per_gpu'] = args.train_batch_size
+    deepspeed_plugin.deepspeed_config['train_batch_size'] = \
+        args.train_batch_size * args.gradient_accumulation_steps * int(os.environ['WORLD_SIZE'])
+    deepspeed_plugin.set_mixed_precision(args.mixed_precision)
+    if args.mixed_precision.lower() == "fp16":
+        deepspeed_plugin.deepspeed_config['fp16']['initial_scale_power'] = 0 # preventing overflow.
+    if args.full_fp16 or args.fp16_master_weights_and_gradients:
+        if args.offload_optimizer_device == "cpu" and args.zero_stage == 2:
+            deepspeed_plugin.deepspeed_config['fp16']['fp16_master_weights_and_grads'] = True
+            print("[DeepSpeed] full fp16 enable.")
+        else:
+            print("[DeepSpeed]full fp16, fp16_master_weights_and_grads currently only supported using ZeRO-Offload with DeepSpeedCPUAdam on ZeRO-2 stage.")
+    
+    if args.offload_optimizer_device is not None:
+        print('[DeepSpeed] start to manually build cpu_adam.')
+        deepspeed.ops.op_builder.CPUAdamBuilder().load()
+        print('[DeepSpeed] building cpu_adam done.')
+
+    return deepspeed_plugin
+
+def prepare_deepspeed_model(args: argparse.Namespace, **models):
+    class DeepSpeedWrapper(torch.nn.Module):
+        def __init__(self, **kw_models) -> None:
+            super().__init__()
+            self.models = torch.nn.ModuleDict()
+
+            for key, model in kw_models.items():
+                if isinstance(model, list):
+                    model = torch.nn.ModuleList(model)
+                assert isinstance(model, torch.nn.Module), f"model must be an instance of torch.nn.Module, but got {key} is {type(model)}"
+                self.models.update(
+                    torch.nn.ModuleDict(
+                        {key: model}
+                    )
+                )
+
+        def get_models(self):
+            return self.models
+
+    ds_model = DeepSpeedWrapper(**models)
+    return ds_model
 
 def prepare_dtype(args: argparse.Namespace):
     weight_dtype = torch.float32
diff --git a/sdxl_train.py b/sdxl_train.py
index 6ffb1bbaf..2f1a5ce65 100644
--- a/sdxl_train.py
+++ b/sdxl_train.py
@@ -391,28 +391,29 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
         text_encoder2.to(weight_dtype)
 
     if args.deepspeed:
-        # Wrapping model for DeepSpeed
-        import deepspeed
-        if args.offload_optimizer_device is not None:
-            accelerator.print('[DeepSpeed] start to manually build cpu_adam.')
-            deepspeed.ops.op_builder.CPUAdamBuilder().load()
-            accelerator.print('[DeepSpeed] building cpu_adam done.')
-
-        class DeepSpeedModel(torch.nn.Module): 
-            def __init__(self, unet, text_encoder) -> None:
-                super().__init__()
-                self.unet = unet
-                self.text_encoders = self.text_encoder = torch.nn.ModuleList(text_encoder)
-                
-            def get_models(self):
-                return self.unet, self.text_encoders
-        text_encoders = [text_encoder1, text_encoder2]
-        ds_model = DeepSpeedModel(unet, text_encoders)
+        training_models_dict = {}
+        if train_unet:
+            training_models_dict["unet"] = unet
+        if train_text_encoder1:
+            text_encoder1.text_model.encoder.layers[-1].requires_grad_(False)
+            text_encoder1.text_model.final_layer_norm.requires_grad_(False)
+            training_models_dict["text_encoder1"] = text_encoder1
+        if train_text_encoder2:
+            training_models_dict["text_encoder2"] = text_encoder2
+        ds_model = train_util.prepare_deepspeed_model(args, **training_models_dict)
         ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(ds_model, optimizer, train_dataloader, lr_scheduler)
-        # Now, ds_model is an instance of DeepSpeedEngine. 
-        unet, text_encoders = ds_model.get_models() # for compatiblility
-        text_encoder1, text_encoder2 = text_encoder = text_encoders
-        training_models = [unet, text_encoder1, text_encoder2]
+        
+        training_models = [] # override training_models
+        if train_unet:
+            unet = ds_model.models["unet"]
+            training_models.append(unet)
+        if train_text_encoder1:
+            text_encoder1 = ds_model.models["text_encoder1"]
+            training_models.append(text_encoder1)
+        if train_text_encoder2:
+            text_encoder2 = ds_model.models["text_encoder2"]
+            training_models.append(text_encoder2)
+
     else: # acceleratorがなんかよろしくやってくれるらしい
         if train_unet:
             unet = accelerator.prepare(unet)
diff --git a/train_db.py b/train_db.py
index c336a1c1c..f188d7bd9 100644
--- a/train_db.py
+++ b/train_db.py
@@ -216,25 +216,20 @@ def train(args):
 
     # acceleratorがなんかよろしくやってくれるらしい
     if args.deepspeed:
-        # wrapping model
-        import deepspeed
-        if args.offload_optimizer_device is not None:
-            accelerator.print('[DeepSpeed] start to manually build cpu_adam.')
-            deepspeed.ops.op_builder.CPUAdamBuilder().load()
-            accelerator.print('[DeepSpeed] building cpu_adam done.')
-        class DeepSpeedModel(torch.nn.Module): 
-            def __init__(self, unet, text_encoder) -> None:
-                super().__init__()
-                self.unet = unet
-                self.text_encoders = self.text_encoder = torch.nn.ModuleList(text_encoder)
-                
-            def get_models(self):
-                return self.unet, self.text_encoders
-        ds_model = DeepSpeedModel(unet, text_encoders)
+        training_models_dict = {}
+        training_models_dict["unet"] = unet
+        if train_text_encoder: training_models_dict["text_encoder"] = text_encoder
+
+        ds_model = train_util.prepare_deepspeed_model(args, **training_models_dict)
         ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(ds_model, optimizer, train_dataloader, lr_scheduler)
-        # Now, ds_model is an instance of DeepSpeedEngine. 
-        unet, text_encoders = ds_model.get_models() # for compatiblility
-        text_encoder = text_encoders
+    
+        training_models = []
+        unet = ds_model.models["unet"]
+        training_models.append(unet)
+        if train_text_encoder:
+            text_encoder = ds_model.models["text_encoder"]
+            training_models.append(text_encoder)
+            
     else:
         if train_text_encoder:
             unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
diff --git a/train_network.py b/train_network.py
index cc445d39a..dfa17eb32 100644
--- a/train_network.py
+++ b/train_network.py
@@ -410,26 +410,22 @@ def train(self, args):
 
         # acceleratorがなんかよろしくやってくれるらしい / accelerator will do something good
         if args.deepspeed:
-            # wrapping model
-            import deepspeed
-            if args.offload_optimizer_device is not None:
-                accelerator.print('[DeepSpeed] start to manually build cpu_adam.')
-                deepspeed.ops.op_builder.CPUAdamBuilder().load()
-                accelerator.print('[DeepSpeed] building cpu_adam done.')
-            class DeepSpeedModel(torch.nn.Module): 
-                def __init__(self, unet, text_encoder, network) -> None:
-                    super().__init__()
-                    self.unet = unet
-                    self.text_encoders = self.text_encoder = torch.nn.ModuleList(text_encoder)
-                    self.network = network
-                    
-                def get_models(self):
-                    return self.unet, self.text_encoders, self.network
-            ds_model = DeepSpeedModel(unet, text_encoders, network)
+            training_models_dict = {}
+            if train_unet: training_models_dict["unet"] = unet
+            if train_text_encoder: training_models_dict["text_encoder"] = text_encoders
+            training_models_dict["network"] = network
+
+            ds_model = train_util.prepare_deepspeed_model(args, **training_models_dict)
             ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(ds_model, optimizer, train_dataloader, lr_scheduler)
-            # Now, ds_model is an instance of DeepSpeedEngine. 
-            unet, text_encoders, network = ds_model.get_models() # for compatiblility
-            text_encoder = text_encoders
+            
+            if train_unet: unet = ds_model.models["unet"]
+            if train_text_encoder:
+                text_encoder = ds_model.models["text_encoder"]
+                if len(ds_model.models["text_encoder"]) > 1:
+                    text_encoders = text_encoder
+                else:
+                    text_encoders = [text_encoder]
+
         else:
             if train_unet:
                 unet = accelerator.prepare(unet)

From 577e9913ca241a56631dbca924e17d9012bde116 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Mon, 26 Feb 2024 20:01:25 +0900
Subject: [PATCH 12/69] add some new dataset settings

---
 README.md              | 161 ++++++++++++++++++++++++++---------------
 library/config_util.py |   6 ++
 library/train_util.py  |  62 +++++++++++++++-
 train_network.py       |   5 ++
 4 files changed, 175 insertions(+), 59 deletions(-)

diff --git a/README.md b/README.md
index e635e5aed..e1b6a26c3 100644
--- a/README.md
+++ b/README.md
@@ -249,6 +249,109 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
 
 ## Change History
 
+### Working in progress
+
+- `train_network.py` and `sdxl_train_network.py` are modified to record some dataset settings in the metadata of the trained model (`caption_prefix`, `caption_suffix`, `keep_tokens_separator`, `secondary_separator`, `enable_wildcard`).
+- Some features are added to the dataset subset settings.
+  - `secondary_separator` is added to specify the tag separator that is not the target of shuffling or dropping. 
+    - Specify `secondary_separator=";;;"`. When you specify `secondary_separator`, the part is not shuffled or dropped. See the example below.
+  - `enable_wildcard` is added. When set to `true`, the wildcard notation `{aaa|bbb|ccc}` can be used. See the example below.
+  - `keep_tokens_separator` is updated to be used twice in the caption. When you specify `keep_tokens_separator="|||"`, the part divided by the second `|||` is not shuffled or dropped and remains at the end.
+  - The existing features `caption_prefix` and `caption_suffix` can be used together. `caption_prefix` and `caption_suffix` are processed first, and then `enable_wildcard`, `keep_tokens_separator`, shuffling and dropping, and `secondary_separator` are processed in order.
+  - The examples are [shown below](#example-of-dataset-settings--データセット設定の記述例).
+
+- `train_network.py` および `sdxl_train_network.py` で、学習したモデルのメタデータに一部のデータセット設定が記録されるよう修正しました（`caption_prefix`、`caption_suffix`、`keep_tokens_separator`、`secondary_separator`、`enable_wildcard`）。
+- データセットのサブセット設定にいくつかの機能を追加しました。
+  - シャッフルの対象とならないタグ分割識別子の指定 `secondary_separator` を追加しました。`secondary_separator=";;;"` のように指定します。`secondary_separator` で区切ることで、その部分はシャッフル、drop 時にまとめて扱われます。詳しくは記述例をご覧ください。
+  - `enable_wildcard` を追加しました。`true` にするとワイルドカード記法 `{aaa|bbb|ccc}` が使えます。詳しくは記述例をご覧ください。
+  - `keep_tokens_separator` をキャプション内に 2 つ使えるようにしました。たとえば `keep_tokens_separator="|||"` と指定したとき、`1girl, hatsune miku, vocaloid ||| stage, mic ||| best quality, rating: general` とキャプションを指定すると、二番目の `|||` で分割された部分はシャッフル、drop されず末尾に残ります。
+  - 既存の機能 `caption_prefix` と `caption_suffix` とあわせて使えます。`caption_prefix` と `caption_suffix` は一番最初に処理され、その後、ワイルドカード、`keep_tokens_separator`、シャッフルおよび drop、`secondary_separator` の順に処理されます。
+
+#### Example of dataset settings / データセット設定の記述例:
+
+```toml
+[general]
+flip_aug = true
+color_aug = false
+resolution = [1024, 1024]
+
+[[datasets]]
+batch_size = 6
+enable_bucket = true
+bucket_no_upscale = true
+caption_extension = ".txt"
+keep_tokens_separator= "|||"
+shuffle_caption = true
+caption_tag_dropout_rate = 0.1
+secondary_separator = ";;;" # subset 側に書くこともできます / can be written in the subset side
+enable_wildcard = true # 同上 / same as above
+
+  [[datasets.subsets]]
+  image_dir = "/path/to/image_dir"
+  num_repeats = 1
+
+  # ||| の前後はカンマは不要です（自動的に追加されます） / No comma is required before and after ||| (it is added automatically)
+  caption_prefix = "1girl, hatsune miku, vocaloid |||" 
+  
+  # ||| の後はシャッフル、drop されず残ります / After |||, it is not shuffled or dropped and remains
+  # 単純に文字列として連結されるので、カンマなどは自分で入れる必要があります / It is simply concatenated as a string, so you need to put commas yourself
+  caption_suffix = ", anime screencap ||| masterpiece, rating: general"
+```
+
+#### Example of caption, secondary_separator notation: `secondary_separator = ";;;"`
+
+```txt
+1girl, hatsune miku, vocaloid, upper body, looking at viewer, sky;;;cloud;;;day, outdoors
+```
+The part `sky;;;cloud;;;day` is replaced with `sky,cloud,day` without shuffling or dropping. When shuffling and dropping are enabled, it is processed as a whole (as one tag). For example, it becomes `vocaloid, 1girl, upper body, sky,cloud,day, outdoors, hatsune miku` (shuffled) or `vocaloid, 1girl, outdoors, looking at viewer, upper body, hatsune miku` (dropped).
+
+#### Example of caption, enable_wildcard notation: `enable_wildcard = true`
+
+```txt
+1girl, hatsune miku, vocaloid, upper body, looking at viewer, {simple|white} background
+```
+`simple` or `white` is randomly selected, and it becomes `simple background` or `white background`.
+
+```txt
+1girl, hatsune miku, vocaloid, {{retro style}}
+```
+If you want to include `{` or `}` in the tag string, double them like `{{` or `}}` (in this example, the actual caption used for training is `{retro style}`).
+
+#### Example of caption, `keep_tokens_separator` notation: `keep_tokens_separator = "|||"`
+
+```txt
+1girl, hatsune miku, vocaloid ||| stage, microphone, white shirt, smile ||| best quality, rating: general
+```
+It becomes `1girl, hatsune miku, vocaloid, microphone, stage, white shirt, best quality, rating: general` or `1girl, hatsune miku, vocaloid, white shirt, smile, stage, microphone, best quality, rating: general` etc.
+
+
+#### キャプション記述例、secondary_separator 記法：`secondary_separator = ";;;"` の場合
+
+```txt
+1girl, hatsune miku, vocaloid, upper body, looking at viewer, sky;;;cloud;;;day, outdoors
+```
+`sky;;;cloud;;;day` の部分はシャッフル、drop されず `sky,cloud,day` に置換されます。シャッフル、drop が有効な場合、まとめて（一つのタグとして）処理されます。つまり `vocaloid, 1girl, upper body, sky,cloud,day, outdoors, hatsune miku` （シャッフル）や `vocaloid, 1girl, outdoors, looking at viewer, upper body, hatsune miku` （drop されたケース）などになります。
+
+#### キャプション記述例、ワイルドカード記法： `enable_wildcard = true` の場合
+
+```txt
+1girl, hatsune miku, vocaloid, upper body, looking at viewer, {simple|white} background
+```
+ランダムに `simple` または `white` が選ばれ、`simple background` または `white background` になります。
+
+```txt
+1girl, hatsune miku, vocaloid, {{retro style}}
+```
+タグ文字列に `{` や `}` そのものを含めたい場合は `{{` や `}}` のように二つ重ねてください（この例では実際に学習に用いられるキャプションは `{retro style}` になります）。
+
+#### キャプション記述例、`keep_tokens_separator` 記法： `keep_tokens_separator = "|||"` の場合
+
+```txt
+1girl, hatsune miku, vocaloid ||| stage, microphone, white shirt, smile ||| best quality, rating: general
+```
+`1girl, hatsune miku, vocaloid, microphone, stage, white shirt, best quality, rating: general` や `1girl, hatsune miku, vocaloid, white shirt, smile, stage, microphone, best quality, rating: general` などになります。
+
+
 ### Feb 24, 2024 / 2024/2/24: v0.8.4
 
 - The log output has been improved. PR [#905](https://github.com/kohya-ss/sd-scripts/pull/905) Thanks to shirayu!
@@ -304,64 +407,6 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
 - 複数 GPU での学習時に `network_multiplier` を指定するとクラッシュする不具合が修正されました。 PR [#1084](https://github.com/kohya-ss/sd-scripts/pull/1084) fireicewolf 氏に感謝します。
 - ControlNet-LLLite の学習がエラーになる不具合を修正しました。 
 
-### Jan 23, 2024 / 2024/1/23: v0.8.2
-
-- [Experimental] The `--fp8_base` option is added to the training scripts for LoRA etc. The base model (U-Net, and Text Encoder when training modules for Text Encoder) can be trained with fp8. PR [#1057](https://github.com/kohya-ss/sd-scripts/pull/1057) Thanks to KohakuBlueleaf!
-  - Please specify `--fp8_base` in `train_network.py` or `sdxl_train_network.py`.
-  - PyTorch 2.1 or later is required.
-  - If you use xformers with PyTorch 2.1, please see [xformers repository](https://github.com/facebookresearch/xformers) and install the appropriate version according to your CUDA version.
-  - The sample image generation during training consumes a lot of memory. It is recommended to turn it off.
-
-- [Experimental] The network multiplier can be specified for each dataset in the training scripts for LoRA etc.
-  - This is an experimental option and may be removed or changed in the future.
-  - For example, if you train with state A as `1.0` and state B as `-1.0`, you may be able to generate by switching between state A and B depending on the LoRA application rate.
-  - Also, if you prepare five states and train them as `0.2`, `0.4`, `0.6`, `0.8`, and `1.0`, you may be able to generate by switching the states smoothly depending on the application rate.
-  - Please specify `network_multiplier` in `[[datasets]]` in `.toml` file.
-- Some options are added to `networks/extract_lora_from_models.py` to reduce the memory usage.
-  - `--load_precision` option can be used to specify the precision when loading the model. If the model is saved in fp16, you can reduce the memory usage by specifying `--load_precision fp16` without losing precision.
-  - `--load_original_model_to` option can be used to specify the device to load the original model. `--load_tuned_model_to` option can be used to specify the device to load the derived model. The default is `cpu` for both options, but you can specify `cuda` etc. You can reduce the memory usage by loading one of them to GPU. This option is available only for SDXL.
-
-- The gradient synchronization in LoRA training with multi-GPU is improved. PR [#1064](https://github.com/kohya-ss/sd-scripts/pull/1064) Thanks to KohakuBlueleaf!
-- The code for Intel IPEX support is improved. PR [#1060](https://github.com/kohya-ss/sd-scripts/pull/1060) Thanks to akx!
-- Fixed a bug in multi-GPU Textual Inversion training.
-
-- （実験的）　LoRA等の学習スクリプトで、ベースモデル（U-Net、および Text Encoder のモジュール学習時は Text Encoder も）の重みを fp8 にして学習するオプションが追加されました。 PR [#1057](https://github.com/kohya-ss/sd-scripts/pull/1057) KohakuBlueleaf 氏に感謝します。
-  - `train_network.py` または `sdxl_train_network.py` で `--fp8_base` を指定してください。
-  - PyTorch 2.1 以降が必要です。
-  - PyTorch 2.1 で xformers を使用する場合は、[xformers のリポジトリ](https://github.com/facebookresearch/xformers) を参照し、CUDA バージョンに応じて適切なバージョンをインストールしてください。
-  - 学習中のサンプル画像生成はメモリを大量に消費するため、オフにすることをお勧めします。
-- (実験的)　LoRA 等の学習で、データセットごとに異なるネットワーク適用率を指定できるようになりました。 
-  - 実験的オプションのため、将来的に削除または仕様変更される可能性があります。
-  - たとえば状態 A を `1.0`、状態 B を `-1.0` として学習すると、LoRA の適用率に応じて状態 A と B を切り替えつつ生成できるかもしれません。
-  - また、五段階の状態を用意し、それぞれ `0.2`、`0.4`、`0.6`、`0.8`、`1.0` として学習すると、適用率でなめらかに状態を切り替えて生成できるかもしれません。 
-  - `.toml` ファイルで `[[datasets]]` に `network_multiplier` を指定してください。
-- `networks/extract_lora_from_models.py` に使用メモリ量を削減するいくつかのオプションを追加しました。 
-  - `--load_precision` で読み込み時の精度を指定できます。モデルが fp16 で保存されている場合は `--load_precision fp16` を指定して精度を変えずにメモリ量を削減できます。
-  - `--load_original_model_to` で元モデルを読み込むデバイスを、`--load_tuned_model_to` で派生モデルを読み込むデバイスを指定できます。デフォルトは両方とも `cpu` ですがそれぞれ `cuda` 等を指定できます。片方を GPU に読み込むことでメモリ量を削減できます。SDXL の場合のみ有効です。
-- マルチ GPU での LoRA 等の学習時に勾配の同期が改善されました。 PR [#1064](https://github.com/kohya-ss/sd-scripts/pull/1064) KohakuBlueleaf 氏に感謝します。
-- Intel IPEX サポートのコードが改善されました。PR [#1060](https://github.com/kohya-ss/sd-scripts/pull/1060) akx 氏に感謝します。
-- マルチ GPU での Textual Inversion 学習の不具合を修正しました。
-
-- `.toml` example for network multiplier / ネットワーク適用率の `.toml` の記述例
-
-```toml
-[general]
-[[datasets]]
-resolution = 512
-batch_size = 8
-network_multiplier = 1.0
-
-... subset settings ...
-
-[[datasets]]
-resolution = 512
-batch_size = 8
-network_multiplier = -1.0
-
-... subset settings ...
-```
-
-
 Please read [Releases](https://github.com/kohya-ss/sd-scripts/releases) for recent updates.
 最近の更新情報は [Release](https://github.com/kohya-ss/sd-scripts/releases) をご覧ください。
 
diff --git a/library/config_util.py b/library/config_util.py
index fc4b36175..eb652ecf3 100644
--- a/library/config_util.py
+++ b/library/config_util.py
@@ -60,6 +60,8 @@ class BaseSubsetParams:
     caption_separator: str = (",",)
     keep_tokens: int = 0
     keep_tokens_separator: str = (None,)
+    secondary_separator: Optional[str] = None
+    enable_wildcard: bool = False
     color_aug: bool = False
     flip_aug: bool = False
     face_crop_aug_range: Optional[Tuple[float, float]] = None
@@ -181,6 +183,8 @@ def __validate_and_convert_scalar_or_twodim(klass, value: Union[float, Sequence]
         "shuffle_caption": bool,
         "keep_tokens": int,
         "keep_tokens_separator": str,
+        "secondary_separator": str,
+        "enable_wildcard": bool,
         "token_warmup_min": int,
         "token_warmup_step": Any(float, int),
         "caption_prefix": str,
@@ -504,6 +508,8 @@ def generate_dataset_group_by_blueprint(dataset_group_blueprint: DatasetGroupBlu
           shuffle_caption: {subset.shuffle_caption}
           keep_tokens: {subset.keep_tokens}
           keep_tokens_separator: {subset.keep_tokens_separator}
+          secondary_separator: {subset.secondary_separator}
+          enable_wildcard: {subset.enable_wildcard}
           caption_dropout_rate: {subset.caption_dropout_rate}
           caption_dropout_every_n_epoches: {subset.caption_dropout_every_n_epochs}
           caption_tag_dropout_rate: {subset.caption_tag_dropout_rate}
diff --git a/library/train_util.py b/library/train_util.py
index d2b69edb5..b71e4edc6 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -364,6 +364,8 @@ def __init__(
         caption_separator: str,
         keep_tokens: int,
         keep_tokens_separator: str,
+        secondary_separator: Optional[str],
+        enable_wildcard: bool,
         color_aug: bool,
         flip_aug: bool,
         face_crop_aug_range: Optional[Tuple[float, float]],
@@ -382,6 +384,8 @@ def __init__(
         self.caption_separator = caption_separator
         self.keep_tokens = keep_tokens
         self.keep_tokens_separator = keep_tokens_separator
+        self.secondary_separator = secondary_separator
+        self.enable_wildcard = enable_wildcard
         self.color_aug = color_aug
         self.flip_aug = flip_aug
         self.face_crop_aug_range = face_crop_aug_range
@@ -410,6 +414,8 @@ def __init__(
         caption_separator: str,
         keep_tokens,
         keep_tokens_separator,
+        secondary_separator,
+        enable_wildcard,
         color_aug,
         flip_aug,
         face_crop_aug_range,
@@ -431,6 +437,8 @@ def __init__(
             caption_separator,
             keep_tokens,
             keep_tokens_separator,
+            secondary_separator,
+            enable_wildcard,
             color_aug,
             flip_aug,
             face_crop_aug_range,
@@ -466,6 +474,8 @@ def __init__(
         caption_separator,
         keep_tokens,
         keep_tokens_separator,
+        secondary_separator,
+        enable_wildcard,
         color_aug,
         flip_aug,
         face_crop_aug_range,
@@ -487,6 +497,8 @@ def __init__(
             caption_separator,
             keep_tokens,
             keep_tokens_separator,
+            secondary_separator,
+            enable_wildcard,
             color_aug,
             flip_aug,
             face_crop_aug_range,
@@ -519,6 +531,8 @@ def __init__(
         caption_separator,
         keep_tokens,
         keep_tokens_separator,
+        secondary_separator,
+        enable_wildcard,
         color_aug,
         flip_aug,
         face_crop_aug_range,
@@ -540,6 +554,8 @@ def __init__(
             caption_separator,
             keep_tokens,
             keep_tokens_separator,
+            secondary_separator,
+            enable_wildcard,
             color_aug,
             flip_aug,
             face_crop_aug_range,
@@ -675,15 +691,41 @@ def process_caption(self, subset: BaseSubset, caption):
         if is_drop_out:
             caption = ""
         else:
+            # process wildcards
+            if subset.enable_wildcard:
+                # wildcard is like '{aaa|bbb|ccc...}'
+                # escape the curly braces like {{ or }}
+                replacer1 = "⦅"
+                replacer2 = "⦆"
+                while replacer1 in caption or replacer2 in caption:
+                    replacer1 += "⦅"
+                    replacer2 += "⦆"
+
+                caption = caption.replace("{{", replacer1).replace("}}", replacer2)
+
+                # replace the wildcard
+                def replace_wildcard(match):
+                    return random.choice(match.group(1).split("|"))
+
+                caption = re.sub(r"\{([^}]+)\}", replace_wildcard, caption)
+
+                # unescape the curly braces
+                caption = caption.replace(replacer1, "{").replace(replacer2, "}")
+
             if subset.shuffle_caption or subset.token_warmup_step > 0 or subset.caption_tag_dropout_rate > 0:
                 fixed_tokens = []
                 flex_tokens = []
+                fixed_suffix_tokens = []
                 if (
                     hasattr(subset, "keep_tokens_separator")
                     and subset.keep_tokens_separator
                     and subset.keep_tokens_separator in caption
                 ):
                     fixed_part, flex_part = caption.split(subset.keep_tokens_separator, 1)
+                    if subset.keep_tokens_separator in flex_part:
+                        flex_part, fixed_suffix_part = flex_part.split(subset.keep_tokens_separator, 1)
+                        fixed_suffix_tokens = [t.strip() for t in fixed_suffix_part.split(subset.caption_separator) if t.strip()]
+
                     fixed_tokens = [t.strip() for t in fixed_part.split(subset.caption_separator) if t.strip()]
                     flex_tokens = [t.strip() for t in flex_part.split(subset.caption_separator) if t.strip()]
                 else:
@@ -718,7 +760,11 @@ def dropout_tags(tokens):
 
                 flex_tokens = dropout_tags(flex_tokens)
 
-                caption = ", ".join(fixed_tokens + flex_tokens)
+                caption = ", ".join(fixed_tokens + flex_tokens + fixed_suffix_tokens)
+
+            # process secondary separator
+            if subset.secondary_separator:
+                caption = caption.replace(subset.secondary_separator, subset.caption_separator)
 
             # textual inversion対応
             for str_from, str_to in self.replacements.items():
@@ -1774,6 +1820,8 @@ def __init__(
                 subset.caption_separator,
                 subset.keep_tokens,
                 subset.keep_tokens_separator,
+                subset.secondary_separator,
+                subset.enable_wildcard,
                 subset.color_aug,
                 subset.flip_aug,
                 subset.face_crop_aug_range,
@@ -3284,6 +3332,18 @@ def add_dataset_arguments(
         help="A custom separator to divide the caption into fixed and flexible parts. Tokens before this separator will not be shuffled. If not specified, '--keep_tokens' will be used to determine the fixed number of tokens."
         + " / captionを固定部分と可変部分に分けるためのカスタム区切り文字。この区切り文字より前のトークンはシャッフルされない。指定しない場合、'--keep_tokens'が固定部分のトークン数として使用される。",
     )
+    parser.add_argument(
+        "--secondary_separator",
+        type=str,
+        default=None,
+        help="a secondary separator for caption. This separator is replaced to caption_separator after dropping/shuffling caption"
+        + " / captionのセカンダリ区切り文字。この区切り文字はcaptionのドロップやシャッフル後にcaption_separatorに置き換えられる",
+    )
+    parser.add_argument(
+        "--enable_wildcard",
+        action="store_true",
+        help="enable wildcard for caption (e.g. '{image|picture|rendition}') / captionのワイルドカードを有効にする（例：'{image|picture|rendition}'）",
+    )
     parser.add_argument(
         "--caption_prefix",
         type=str,
diff --git a/train_network.py b/train_network.py
index e0fa69458..e5b26d8a2 100644
--- a/train_network.py
+++ b/train_network.py
@@ -564,6 +564,11 @@ def train(self, args):
                         "random_crop": bool(subset.random_crop),
                         "shuffle_caption": bool(subset.shuffle_caption),
                         "keep_tokens": subset.keep_tokens,
+                        "keep_tokens_separator": subset.keep_tokens_separator,
+                        "secondary_separator": subset.secondary_separator,
+                        "enable_wildcard": bool(subset.enable_wildcard),
+                        "caption_prefix": subset.caption_prefix,
+                        "caption_suffix": subset.caption_suffix,
                     }
 
                     image_dir_or_metadata_file = None

From f2c727fc8cadf0971c24fdb42c8684032e7e6f80 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Mon, 26 Feb 2024 23:19:58 +0900
Subject: [PATCH 13/69] add minimal impl  for masked loss

---
 library/config_util.py | 38 +++++++++++++++++++++++++-------------
 library/train_util.py  |  3 +++
 train_network.py       | 18 +++++++++++++++++-
 3 files changed, 45 insertions(+), 14 deletions(-)

diff --git a/library/config_util.py b/library/config_util.py
index eb652ecf3..edc6a5385 100644
--- a/library/config_util.py
+++ b/library/config_util.py
@@ -41,12 +41,17 @@
     DatasetGroup,
 )
 from .utils import setup_logging
+
 setup_logging()
 import logging
+
 logger = logging.getLogger(__name__)
 
+
 def add_config_arguments(parser: argparse.ArgumentParser):
-    parser.add_argument("--dataset_config", type=Path, default=None, help="config file for detail settings / 詳細な設定用の設定ファイル")
+    parser.add_argument(
+        "--dataset_config", type=Path, default=None, help="config file for detail settings / 詳細な設定用の設定ファイル"
+    )
 
 
 # TODO: inherit Params class in Subset, Dataset
@@ -248,9 +253,10 @@ def __validate_and_convert_scalar_or_twodim(klass, value: Union[float, Sequence]
     }
 
     def __init__(self, support_dreambooth: bool, support_finetuning: bool, support_controlnet: bool, support_dropout: bool) -> None:
-        assert (
-            support_dreambooth or support_finetuning or support_controlnet
-        ), "Neither DreamBooth mode nor fine tuning mode specified. Please specify one mode or more. / DreamBooth モードか fine tuning モードのどちらも指定されていません。1つ以上指定してください。"
+        assert support_dreambooth or support_finetuning or support_controlnet, (
+            "Neither DreamBooth mode nor fine tuning mode nor controlnet mode specified. Please specify one mode or more."
+            + " / DreamBooth モードか fine tuning モードか controlnet モードのどれも指定されていません。1つ以上指定してください。"
+        )
 
         self.db_subset_schema = self.__merge_dict(
             self.SUBSET_ASCENDABLE_SCHEMA,
@@ -362,7 +368,9 @@ def sanitize_argparse_namespace(self, argparse_namespace: argparse.Namespace) ->
             return self.argparse_config_validator(argparse_namespace)
         except MultipleInvalid:
             # XXX: this should be a bug
-            logger.error("Invalid cmdline parsed arguments. This should be a bug. / コマンドラインのパース結果が正しくないようです。プログラムのバグの可能性が高いです。")
+            logger.error(
+                "Invalid cmdline parsed arguments. This should be a bug. / コマンドラインのパース結果が正しくないようです。プログラムのバグの可能性が高いです。"
+            )
             raise
 
     # NOTE: value would be overwritten by latter dict if there is already the same key
@@ -547,11 +555,11 @@ def generate_dataset_group_by_blueprint(dataset_group_blueprint: DatasetGroupBlu
                     "    ",
                 )
 
-    logger.info(f'{info}')
+    logger.info(f"{info}")
 
     # make buckets first because it determines the length of dataset
     # and set the same seed for all datasets
-    seed = random.randint(0, 2**31) # actual seed is seed + epoch_no
+    seed = random.randint(0, 2**31)  # actual seed is seed + epoch_no
     for i, dataset in enumerate(datasets):
         logger.info(f"[Dataset {i}]")
         dataset.make_buckets()
@@ -638,13 +646,17 @@ def load_user_config(file: str) -> dict:
             with open(file, "r") as f:
                 config = json.load(f)
         except Exception:
-            logger.error(f"Error on parsing JSON config file. Please check the format. / JSON 形式の設定ファイルの読み込みに失敗しました。文法が正しいか確認してください。: {file}")
+            logger.error(
+                f"Error on parsing JSON config file. Please check the format. / JSON 形式の設定ファイルの読み込みに失敗しました。文法が正しいか確認してください。: {file}"
+            )
             raise
     elif file.name.lower().endswith(".toml"):
         try:
             config = toml.load(file)
         except Exception:
-            logger.error(f"Error on parsing TOML config file. Please check the format. / TOML 形式の設定ファイルの読み込みに失敗しました。文法が正しいか確認してください。: {file}")
+            logger.error(
+                f"Error on parsing TOML config file. Please check the format. / TOML 形式の設定ファイルの読み込みに失敗しました。文法が正しいか確認してください。: {file}"
+            )
             raise
     else:
         raise ValueError(f"not supported config file format / 対応していない設定ファイルの形式です: {file}")
@@ -671,13 +683,13 @@ def load_user_config(file: str) -> dict:
     train_util.prepare_dataset_args(argparse_namespace, config_args.support_finetuning)
 
     logger.info("[argparse_namespace]")
-    logger.info(f'{vars(argparse_namespace)}')
+    logger.info(f"{vars(argparse_namespace)}")
 
     user_config = load_user_config(config_args.dataset_config)
 
     logger.info("")
     logger.info("[user_config]")
-    logger.info(f'{user_config}')
+    logger.info(f"{user_config}")
 
     sanitizer = ConfigSanitizer(
         config_args.support_dreambooth, config_args.support_finetuning, config_args.support_controlnet, config_args.support_dropout
@@ -686,10 +698,10 @@ def load_user_config(file: str) -> dict:
 
     logger.info("")
     logger.info("[sanitized_user_config]")
-    logger.info(f'{sanitized_user_config}')
+    logger.info(f"{sanitized_user_config}")
 
     blueprint = BlueprintGenerator(sanitizer).generate(user_config, argparse_namespace)
 
     logger.info("")
     logger.info("[blueprint]")
-    logger.info(f'{blueprint}')
+    logger.info(f"{blueprint}")
diff --git a/library/train_util.py b/library/train_util.py
index b71e4edc6..7fe5bc56e 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -1810,6 +1810,9 @@ def __init__(
 
         db_subsets = []
         for subset in subsets:
+            assert (
+                not subset.random_crop
+            ), "random_crop is not supported in ControlNetDataset / random_cropはControlNetDatasetではサポートされていません"
             db_subset = DreamBoothSubset(
                 subset.image_dir,
                 False,
diff --git a/train_network.py b/train_network.py
index e5b26d8a2..e3ce7bd36 100644
--- a/train_network.py
+++ b/train_network.py
@@ -13,6 +13,7 @@
 
 import torch
 from library.device_utils import init_ipex, clean_memory_on_device
+
 init_ipex()
 
 from torch.nn.parallel import DistributedDataParallel as DDP
@@ -157,7 +158,7 @@ def train(self, args):
 
         # データセットを準備する
         if args.dataset_class is None:
-            blueprint_generator = BlueprintGenerator(ConfigSanitizer(True, True, False, True))
+            blueprint_generator = BlueprintGenerator(ConfigSanitizer(True, True, args.masked_loss, True))
             if use_user_config:
                 logger.info(f"Loading dataset config from {args.dataset_config}")
                 user_config = config_util.load_user_config(args.dataset_config)
@@ -834,6 +835,16 @@ def remove_model(old_ckpt_name):
                         target = noise
 
                     loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none")
+
+                    if args.masked_loss:
+                        # mask image is -1 to 1. we need to convert it to 0 to 1
+                        mask_image = batch["conditioning_images"].to(dtype=weight_dtype)[:, 0].unsqueeze(1)  # use R channel
+
+                        # resize to the same size as the loss
+                        mask_image = torch.nn.functional.interpolate(mask_image, size=loss.shape[2:], mode="area")
+                        mask_image = mask_image / 2 + 0.5
+                        loss = loss * mask_image
+
                     loss = loss.mean([1, 2, 3])
 
                     loss_weights = batch["loss_weights"]  # 各sampleごとのweight
@@ -1050,6 +1061,11 @@ def setup_parser() -> argparse.ArgumentParser:
         action="store_true",
         help="do not use fp16/bf16 VAE in mixed precision (use float VAE) / mixed precisionでも fp16/bf16 VAEを使わずfloat VAEを使う",
     )
+    parser.add_argument(
+        "--masked_loss",
+        action="store_true",
+        help="apply mask for caclulating loss. conditioning_data_dir is required for dataset. / 損失計算時にマスクを適用する。datasetにはconditioning_data_dirが必要",
+    )
     return parser
 
 

From 175193623b39027ffcfe0c0ae250dbce564ed6ef Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Mon, 26 Feb 2024 23:29:41 +0900
Subject: [PATCH 14/69] update readme

---
 README.md | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/README.md b/README.md
index e1b6a26c3..9cc79cc09 100644
--- a/README.md
+++ b/README.md
@@ -249,6 +249,13 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
 
 ## Change History
 
+### Masked loss
+
+`train_network.py` and `sdxl_train_network.py` now support the masked loss. `--masked_loss` option is added. 
+
+ControlNet dataset is used to specify the mask. The mask images should be the RGB images. The pixel value 255 in R channel is treated as the mask (the loss is calculated only for the pixels with the mask), and 0 is treated as the non-mask. See details for the dataset specification in the [LLLite documentation](./docs/train_lllite_README.md#preparing-the-dataset).
+
+
 ### Working in progress
 
 - `train_network.py` and `sdxl_train_network.py` are modified to record some dataset settings in the metadata of the trained model (`caption_prefix`, `caption_suffix`, `keep_tokens_separator`, `secondary_separator`, `enable_wildcard`).

From 4a5546d40e6de5789be78dd16373d2b820b8754e Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Mon, 26 Feb 2024 23:39:56 +0900
Subject: [PATCH 15/69] fix typo

---
 train_network.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/train_network.py b/train_network.py
index e3ce7bd36..f5617986c 100644
--- a/train_network.py
+++ b/train_network.py
@@ -1064,7 +1064,7 @@ def setup_parser() -> argparse.ArgumentParser:
     parser.add_argument(
         "--masked_loss",
         action="store_true",
-        help="apply mask for caclulating loss. conditioning_data_dir is required for dataset. / 損失計算時にマスクを適用する。datasetにはconditioning_data_dirが必要",
+        help="apply mask for calculating loss. conditioning_data_dir is required for dataset. / 損失計算時にマスクを適用する。datasetにはconditioning_data_dirが必要",
     )
     return parser
 

From e3ccf8fbf73a0f728fc167a20b1e0648a3604f41 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Tue, 27 Feb 2024 21:30:46 +0900
Subject: [PATCH 16/69] make deepspeed_utils

---
 fine_tune.py               |  35 +++++-----
 library/deepspeed_utils.py | 139 +++++++++++++++++++++++++++++++++++++
 library/train_util.py      | 110 ++---------------------------
 sdxl_train.py              |  64 ++++++++---------
 train_db.py                |  39 ++++++-----
 train_network.py           |  51 +++++++-------
 6 files changed, 238 insertions(+), 200 deletions(-)
 create mode 100644 library/deepspeed_utils.py

diff --git a/fine_tune.py b/fine_tune.py
index c5e97d267..b018a933d 100644
--- a/fine_tune.py
+++ b/fine_tune.py
@@ -10,7 +10,9 @@
 from tqdm import tqdm
 
 import torch
+from library import deepspeed_utils
 from library.device_utils import init_ipex, clean_memory_on_device
+
 init_ipex()
 
 from accelerate.utils import set_seed
@@ -42,6 +44,7 @@
 def train(args):
     train_util.verify_training_args(args)
     train_util.prepare_dataset_args(args, True)
+    deepspeed_utils.prepare_deepspeed_args(args)
     setup_logging(args, reset=True)
 
     cache_latents = args.cache_latents
@@ -219,7 +222,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
         batch_size=1,
         shuffle=True,
         collate_fn=collator,
-        num_workers=n_workers if not args.deepspeed else 1, # To avoid RuntimeError: DataLoader worker exited unexpectedly with exit code 1.
+        num_workers=n_workers,
         persistent_workers=args.persistent_data_loader_workers,
     )
 
@@ -231,7 +234,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
         accelerator.print(
             f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
         )
-        
+
     # データセット側にも学習ステップを送信
     train_dataset_group.set_max_train_steps(args.max_train_steps)
 
@@ -248,21 +251,16 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
         text_encoder.to(weight_dtype)
 
     if args.deepspeed:
-        training_models_dict = {}
-        training_models_dict["unet"] = unet
-        if args.train_text_encoder: training_models_dict["text_encoder"] = text_encoder
-
-        ds_model = train_util.prepare_deepspeed_model(args, **training_models_dict)
-        ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(ds_model, optimizer, train_dataloader, lr_scheduler)
-    
-        training_models = []
-        unet = ds_model.models["unet"]
-        training_models.append(unet)
         if args.train_text_encoder:
-            text_encoder = ds_model.models["text_encoder"]
-            training_models.append(text_encoder)
-            
-    else: # acceleratorがなんかよろしくやってくれるらしい
+            ds_model = deepspeed_utils.prepare_deepspeed_model(args, unet=unet, text_encoder=text_encoder)
+        else:
+            ds_model = deepspeed_utils.prepare_deepspeed_model(args, unet=unet)
+        ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+            ds_model, optimizer, train_dataloader, lr_scheduler
+        )
+        training_models = [ds_model]
+    else:
+        # acceleratorがなんかよろしくやってくれるらしい
         if args.train_text_encoder:
             unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
                 unet, text_encoder, optimizer, train_dataloader, lr_scheduler
@@ -327,13 +325,13 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
 
         for step, batch in enumerate(train_dataloader):
             current_step.value = global_step
-            with accelerator.accumulate(training_models[0]):  # 複数モデルに対応していない模様だがとりあえずこうしておく
+            with accelerator.accumulate(*training_models):
                 with torch.no_grad():
                     if "latents" in batch and batch["latents"] is not None:
                         latents = batch["latents"].to(accelerator.device)  # .to(dtype=weight_dtype)
                     else:
                         # latentに変換
-                        latents = vae.encode(batch["images"].to(dtype=weight_dtype)).latent_dist.sample()
+                        latents = vae.encode(batch["images"].to(dtype=vae_dtype)).latent_dist.sample().to(weight_dtype)
                     latents = latents * 0.18215
                 b_size = latents.shape[0]
 
@@ -493,6 +491,7 @@ def setup_parser() -> argparse.ArgumentParser:
     train_util.add_sd_models_arguments(parser)
     train_util.add_dataset_arguments(parser, False, True, True)
     train_util.add_training_arguments(parser, False)
+    deepspeed_utils.add_deepspeed_arguments(parser)
     train_util.add_sd_saving_arguments(parser)
     train_util.add_optimizer_arguments(parser)
     config_util.add_config_arguments(parser)
diff --git a/library/deepspeed_utils.py b/library/deepspeed_utils.py
new file mode 100644
index 000000000..99a7b2b3b
--- /dev/null
+++ b/library/deepspeed_utils.py
@@ -0,0 +1,139 @@
+import os
+import argparse
+import torch
+from accelerate import DeepSpeedPlugin, Accelerator
+
+from .utils import setup_logging
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+def add_deepspeed_arguments(parser: argparse.ArgumentParser):
+    # DeepSpeed Arguments. https://huggingface.co/docs/accelerate/usage_guides/deepspeed
+    parser.add_argument("--deepspeed", action="store_true", help="enable deepspeed training")
+    parser.add_argument("--zero_stage", type=int, default=2, choices=[0, 1, 2, 3], help="Possible options are 0,1,2,3.")
+    parser.add_argument(
+        "--offload_optimizer_device",
+        type=str,
+        default=None,
+        choices=[None, "cpu", "nvme"],
+        help="Possible options are none|cpu|nvme. Only applicable with ZeRO Stages 2 and 3.",
+    )
+    parser.add_argument(
+        "--offload_optimizer_nvme_path",
+        type=str,
+        default=None,
+        help="Possible options are /nvme|/local_nvme. Only applicable with ZeRO Stage 3.",
+    )
+    parser.add_argument(
+        "--offload_param_device",
+        type=str,
+        default=None,
+        choices=[None, "cpu", "nvme"],
+        help="Possible options are none|cpu|nvme. Only applicable with ZeRO Stage 3.",
+    )
+    parser.add_argument(
+        "--offload_param_nvme_path",
+        type=str,
+        default=None,
+        help="Possible options are /nvme|/local_nvme. Only applicable with ZeRO Stage 3.",
+    )
+    parser.add_argument(
+        "--zero3_init_flag",
+        action="store_true",
+        help="Flag to indicate whether to enable `deepspeed.zero.Init` for constructing massive models."
+        "Only applicable with ZeRO Stage-3.",
+    )
+    parser.add_argument(
+        "--zero3_save_16bit_model",
+        action="store_true",
+        help="Flag to indicate whether to save 16-bit model. Only applicable with ZeRO Stage-3.",
+    )
+    parser.add_argument(
+        "--fp16_master_weights_and_gradients",
+        action="store_true",
+        help="fp16_master_and_gradients requires optimizer to support keeping fp16 master and gradients while keeping the optimizer states in fp32.",
+    )
+
+
+def prepare_deepspeed_args(args: argparse.Namespace):
+    if not args.deepspeed:
+        return
+
+    # To avoid RuntimeError: DataLoader worker exited unexpectedly with exit code 1.
+    args.max_data_loader_n_workers = 1
+
+
+def prepare_deepspeed_plugin(args: argparse.Namespace):
+    if not args.deepspeed:
+        return None
+
+    try:
+        import deepspeed
+    except ImportError as e:
+        logger.error(
+            "deepspeed is not installed. please install deepspeed in your environment with following command. DS_BUILD_OPS=0 pip install deepspeed"
+        )
+        exit(1)
+
+    deepspeed_plugin = DeepSpeedPlugin(
+        zero_stage=args.zero_stage,
+        gradient_accumulation_steps=args.gradient_accumulation_steps,
+        gradient_clipping=args.max_grad_norm,
+        offload_optimizer_device=args.offload_optimizer_device,
+        offload_optimizer_nvme_path=args.offload_optimizer_nvme_path,
+        offload_param_device=args.offload_param_device,
+        offload_param_nvme_path=args.offload_param_nvme_path,
+        zero3_init_flag=args.zero3_init_flag,
+        zero3_save_16bit_model=args.zero3_save_16bit_model,
+    )
+    deepspeed_plugin.deepspeed_config["train_micro_batch_size_per_gpu"] = args.train_batch_size
+    deepspeed_plugin.deepspeed_config["train_batch_size"] = (
+        args.train_batch_size * args.gradient_accumulation_steps * int(os.environ["WORLD_SIZE"])
+    )
+    deepspeed_plugin.set_mixed_precision(args.mixed_precision)
+    if args.mixed_precision.lower() == "fp16":
+        deepspeed_plugin.deepspeed_config["fp16"]["initial_scale_power"] = 0  # preventing overflow.
+    if args.full_fp16 or args.fp16_master_weights_and_gradients:
+        if args.offload_optimizer_device == "cpu" and args.zero_stage == 2:
+            deepspeed_plugin.deepspeed_config["fp16"]["fp16_master_weights_and_grads"] = True
+            logger.info("[DeepSpeed] full fp16 enable.")
+        else:
+            logger.info(
+                "[DeepSpeed]full fp16, fp16_master_weights_and_grads currently only supported using ZeRO-Offload with DeepSpeedCPUAdam on ZeRO-2 stage."
+            )
+
+    if args.offload_optimizer_device is not None:
+        logger.info("[DeepSpeed] start to manually build cpu_adam.")
+        deepspeed.ops.op_builder.CPUAdamBuilder().load()
+        logger.info("[DeepSpeed] building cpu_adam done.")
+
+    return deepspeed_plugin
+
+
+# Accelerate library does not support multiple models for deepspeed. So, we need to wrap multiple models into a single model.
+def prepare_deepspeed_model(args: argparse.Namespace, **models):
+    # remove None from models
+    models = {k: v for k, v in models.items() if v is not None}
+
+    class DeepSpeedWrapper(torch.nn.Module):
+        def __init__(self, **kw_models) -> None:
+            super().__init__()
+            self.models = torch.nn.ModuleDict()
+
+            for key, model in kw_models.items():
+                if isinstance(model, list):
+                    model = torch.nn.ModuleList(model)
+                assert isinstance(
+                    model, torch.nn.Module
+                ), f"model must be an instance of torch.nn.Module, but got {key} is {type(model)}"
+                self.models.update(torch.nn.ModuleDict({key: model}))
+
+        def get_models(self):
+            return self.models
+
+    ds_model = DeepSpeedWrapper(**models)
+    return ds_model
diff --git a/library/train_util.py b/library/train_util.py
index 3781dcde8..38e1b458d 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -21,7 +21,6 @@
     Union,
 )
 from accelerate import Accelerator, InitProcessGroupKwargs, DistributedDataParallelKwargs
-from accelerate import DeepSpeedPlugin
 import glob
 import math
 import os
@@ -70,6 +69,7 @@
 import library.model_util as model_util
 import library.huggingface_util as huggingface_util
 import library.sai_model_spec as sai_model_spec
+import library.deepspeed_utils as deepspeed_utils
 from library.utils import setup_logging
 
 setup_logging()
@@ -3243,52 +3243,6 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
             "--prior_loss_weight", type=float, default=1.0, help="loss weight for regularization images / 正則化画像のlossの重み"
         )
 
-    # DeepSpeed Arguments. https://huggingface.co/docs/accelerate/usage_guides/deepspeed
-    parser.add_argument("--deepspeed", action="store_true", help="enable deepspeed training")
-    parser.add_argument(
-        "--zero_stage", 
-        type=int, default=2,
-        choices=[0, 1, 2, 3],
-        help="Possible options are 0,1,2,3."
-    )
-    parser.add_argument(
-        "--offload_optimizer_device", 
-        type=str, default=None,
-        choices=[None, "cpu", "nvme"],
-        help="Possible options are none|cpu|nvme. Only applicable with ZeRO Stages 2 and 3."
-    )
-    parser.add_argument(
-        "--offload_optimizer_nvme_path",
-        type=str, default=None,
-        help="Possible options are /nvme|/local_nvme. Only applicable with ZeRO Stage 3."
-    )
-    parser.add_argument(
-        "--offload_param_device",
-        type=str, default=None,
-        choices=[None, "cpu", "nvme"],
-        help="Possible options are none|cpu|nvme. Only applicable with ZeRO Stage 3."
-    )
-    parser.add_argument(
-        "--offload_param_nvme_path",
-        type=str, default=None,
-        help="Possible options are /nvme|/local_nvme. Only applicable with ZeRO Stage 3."
-    )
-    parser.add_argument(
-        "--zero3_init_flag",
-        action="store_true",
-        help="Flag to indicate whether to enable `deepspeed.zero.Init` for constructing massive models."
-            "Only applicable with ZeRO Stage-3."
-    )
-    parser.add_argument(
-        "--zero3_save_16bit_model",
-        action="store_true",
-        help="Flag to indicate whether to save 16-bit model. Only applicable with ZeRO Stage-3."
-    )
-    parser.add_argument(
-        "--fp16_master_weights_and_gradients",
-        action="store_true",
-        help="fp16_master_and_gradients requires optimizer to support keeping fp16 master and gradients while keeping the optimizer states in fp32."
-    )
 
 def verify_training_args(args: argparse.Namespace):
     r"""
@@ -4090,6 +4044,10 @@ def load_tokenizer(args: argparse.Namespace):
 
 
 def prepare_accelerator(args: argparse.Namespace):
+    """
+    this function also prepares deepspeed plugin
+    """
+
     if args.logging_dir is None:
         logging_dir = None
     else:
@@ -4135,7 +4093,7 @@ def prepare_accelerator(args: argparse.Namespace):
         ),
     )
     kwargs_handlers = list(filter(lambda x: x is not None, kwargs_handlers))
-    deepspeed_plugin = prepare_deepspeed_plugin(args)
+    deepspeed_plugin = deepspeed_utils.prepare_deepspeed_plugin(args)
 
     accelerator = Accelerator(
         gradient_accumulation_steps=args.gradient_accumulation_steps,
@@ -4149,62 +4107,6 @@ def prepare_accelerator(args: argparse.Namespace):
     print("accelerator device:", accelerator.device)
     return accelerator
 
-def prepare_deepspeed_plugin(args: argparse.Namespace):
-    if args.deepspeed is None: return None
-    try:
-        import deepspeed
-    except ImportError as e:
-        print("deepspeed is not installed. please install deepspeed in your environment with following command. DS_BUILD_OPS=0 pip install deepspeed")
-        exit(1)
-
-    deepspeed_plugin = DeepSpeedPlugin(
-        zero_stage=args.zero_stage,
-        gradient_accumulation_steps=args.gradient_accumulation_steps, gradient_clipping=args.max_grad_norm,
-        offload_optimizer_device=args.offload_optimizer_device, offload_optimizer_nvme_path=args.offload_optimizer_nvme_path,
-        offload_param_device=args.offload_param_device, offload_param_nvme_path=args.offload_param_nvme_path,
-        zero3_init_flag=args.zero3_init_flag, zero3_save_16bit_model=args.zero3_save_16bit_model,
-    )
-    deepspeed_plugin.deepspeed_config['train_micro_batch_size_per_gpu'] = args.train_batch_size
-    deepspeed_plugin.deepspeed_config['train_batch_size'] = \
-        args.train_batch_size * args.gradient_accumulation_steps * int(os.environ['WORLD_SIZE'])
-    deepspeed_plugin.set_mixed_precision(args.mixed_precision)
-    if args.mixed_precision.lower() == "fp16":
-        deepspeed_plugin.deepspeed_config['fp16']['initial_scale_power'] = 0 # preventing overflow.
-    if args.full_fp16 or args.fp16_master_weights_and_gradients:
-        if args.offload_optimizer_device == "cpu" and args.zero_stage == 2:
-            deepspeed_plugin.deepspeed_config['fp16']['fp16_master_weights_and_grads'] = True
-            print("[DeepSpeed] full fp16 enable.")
-        else:
-            print("[DeepSpeed]full fp16, fp16_master_weights_and_grads currently only supported using ZeRO-Offload with DeepSpeedCPUAdam on ZeRO-2 stage.")
-    
-    if args.offload_optimizer_device is not None:
-        print('[DeepSpeed] start to manually build cpu_adam.')
-        deepspeed.ops.op_builder.CPUAdamBuilder().load()
-        print('[DeepSpeed] building cpu_adam done.')
-
-    return deepspeed_plugin
-
-def prepare_deepspeed_model(args: argparse.Namespace, **models):
-    class DeepSpeedWrapper(torch.nn.Module):
-        def __init__(self, **kw_models) -> None:
-            super().__init__()
-            self.models = torch.nn.ModuleDict()
-
-            for key, model in kw_models.items():
-                if isinstance(model, list):
-                    model = torch.nn.ModuleList(model)
-                assert isinstance(model, torch.nn.Module), f"model must be an instance of torch.nn.Module, but got {key} is {type(model)}"
-                self.models.update(
-                    torch.nn.ModuleDict(
-                        {key: model}
-                    )
-                )
-
-        def get_models(self):
-            return self.models
-
-    ds_model = DeepSpeedWrapper(**models)
-    return ds_model
 
 def prepare_dtype(args: argparse.Namespace):
     weight_dtype = torch.float32
diff --git a/sdxl_train.py b/sdxl_train.py
index 5e5e9f291..0feb4e367 100644
--- a/sdxl_train.py
+++ b/sdxl_train.py
@@ -11,11 +11,12 @@
 
 import torch
 from library.device_utils import init_ipex, clean_memory_on_device
+
 init_ipex()
 
 from accelerate.utils import set_seed
 from diffusers import DDPMScheduler
-from library import sdxl_model_util
+from library import deepspeed_utils, sdxl_model_util
 
 import library.train_util as train_util
 
@@ -97,6 +98,7 @@ def train(args):
     train_util.verify_training_args(args)
     train_util.prepare_dataset_args(args, True)
     sdxl_train_util.verify_sdxl_training_args(args)
+    deepspeed_utils.prepare_deepspeed_args(args)
     setup_logging(args, reset=True)
 
     assert (
@@ -361,7 +363,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
         batch_size=1,
         shuffle=True,
         collate_fn=collator,
-        num_workers=n_workers if not args.deepspeed else 1, # To avoid RuntimeError: DataLoader worker exited unexpectedly with exit code 1.
+        num_workers=n_workers,
         persistent_workers=args.persistent_data_loader_workers,
     )
 
@@ -398,41 +400,31 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
         text_encoder1.to(weight_dtype)
         text_encoder2.to(weight_dtype)
 
+    # freeze last layer and final_layer_norm in te1 since we use the output of the penultimate layer
+    if train_text_encoder1:
+        text_encoder1.text_model.encoder.layers[-1].requires_grad_(False)
+        text_encoder1.text_model.final_layer_norm.requires_grad_(False)
+
     if args.deepspeed:
-        training_models_dict = {}
-        if train_unet:
-            training_models_dict["unet"] = unet
-        if train_text_encoder1:
-            text_encoder1.text_model.encoder.layers[-1].requires_grad_(False)
-            text_encoder1.text_model.final_layer_norm.requires_grad_(False)
-            training_models_dict["text_encoder1"] = text_encoder1
-        if train_text_encoder2:
-            training_models_dict["text_encoder2"] = text_encoder2
-        ds_model = train_util.prepare_deepspeed_model(args, **training_models_dict)
-        ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(ds_model, optimizer, train_dataloader, lr_scheduler)
-        
-        training_models = [] # override training_models
-        if train_unet:
-            unet = ds_model.models["unet"]
-            training_models.append(unet)
-        if train_text_encoder1:
-            text_encoder1 = ds_model.models["text_encoder1"]
-            training_models.append(text_encoder1)
-        if train_text_encoder2:
-            text_encoder2 = ds_model.models["text_encoder2"]
-            training_models.append(text_encoder2)
+        ds_model = deepspeed_utils.prepare_deepspeed_model(
+            args,
+            unet=unet if train_unet else None,
+            text_encoder1=text_encoder1 if train_text_encoder1 else None,
+            text_encoder2=text_encoder2 if train_text_encoder2 else None,
+        )
+        ds_model = accelerator.prepare(ds_model)
+        training_models = [ds_model]
 
-    else: # acceleratorがなんかよろしくやってくれるらしい
+    else:
+        # acceleratorがなんかよろしくやってくれるらしい
         if train_unet:
             unet = accelerator.prepare(unet)
         if train_text_encoder1:
-            # freeze last layer and final_layer_norm in te1 since we use the output of the penultimate layer
-            text_encoder1.text_model.encoder.layers[-1].requires_grad_(False)
-            text_encoder1.text_model.final_layer_norm.requires_grad_(False)
             text_encoder1 = accelerator.prepare(text_encoder1)
         if train_text_encoder2:
             text_encoder2 = accelerator.prepare(text_encoder2)
-        optimizer, train_dataloader, lr_scheduler = accelerator.prepare(optimizer, train_dataloader, lr_scheduler)
+
+    optimizer, train_dataloader, lr_scheduler = accelerator.prepare(optimizer, train_dataloader, lr_scheduler)
 
     # TextEncoderの出力をキャッシュするときにはCPUへ移動する
     if args.cache_text_encoder_outputs:
@@ -446,8 +438,9 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
         text_encoder2.to(accelerator.device)
 
     # 実験的機能：勾配も含めたfp16学習を行う　PyTorchにパッチを当ててfp16でのgrad scaleを有効にする
-    if args.full_fp16 and not args.deepspeed:
+    if args.full_fp16:
         # During deepseed training, accelerate not handles fp16/bf16|mixed precision directly via scaler. Let deepspeed engine do.
+        # -> But we think it's ok to patch accelerator even if deepspeed is enabled.
         train_util.patch_accelerator_for_fp16_training(accelerator)
 
     # resumeする
@@ -508,10 +501,10 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
         for step, batch in enumerate(train_dataloader):
             current_step.value = global_step
             with accelerator.accumulate(*training_models):
-                with torch.no_grad(): # why this block differ within train_network.py?
-                    if "latents" in batch and batch["latents"] is not None:
-                        latents = batch["latents"].to(accelerator.device).to(dtype=weight_dtype)
-                    else:
+                if "latents" in batch and batch["latents"] is not None:
+                    latents = batch["latents"].to(accelerator.device).to(dtype=weight_dtype)
+                else:
+                    with torch.no_grad():
                         # latentに変換
                         latents = vae.encode(batch["images"].to(vae_dtype)).latent_dist.sample().to(weight_dtype)
 
@@ -519,7 +512,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
                         if torch.any(torch.isnan(latents)):
                             accelerator.print("NaN found in latents, replacing with zeros")
                             latents = torch.nan_to_num(latents, 0, out=latents)
-                    latents = latents * sdxl_model_util.VAE_SCALE_FACTOR
+                latents = latents * sdxl_model_util.VAE_SCALE_FACTOR
 
                 if "text_encoder_outputs1_list" not in batch or batch["text_encoder_outputs1_list"] is None:
                     input_ids1 = batch["input_ids"]
@@ -768,6 +761,7 @@ def setup_parser() -> argparse.ArgumentParser:
     train_util.add_sd_models_arguments(parser)
     train_util.add_dataset_arguments(parser, True, True, True)
     train_util.add_training_arguments(parser, False)
+    deepspeed_utils.add_deepspeed_arguments(parser)
     train_util.add_sd_saving_arguments(parser)
     train_util.add_optimizer_arguments(parser)
     config_util.add_config_arguments(parser)
diff --git a/train_db.py b/train_db.py
index 66a83d1df..ea1cfeb8a 100644
--- a/train_db.py
+++ b/train_db.py
@@ -11,7 +11,9 @@
 from tqdm import tqdm
 
 import torch
+from library import deepspeed_utils
 from library.device_utils import init_ipex, clean_memory_on_device
+
 init_ipex()
 
 from accelerate.utils import set_seed
@@ -46,6 +48,7 @@
 def train(args):
     train_util.verify_training_args(args)
     train_util.prepare_dataset_args(args, False)
+    deepspeed_utils.prepare_deepspeed_args(args)
     setup_logging(args, reset=True)
 
     cache_latents = args.cache_latents
@@ -187,7 +190,7 @@ def train(args):
         batch_size=1,
         shuffle=True,
         collate_fn=collator,
-        num_workers=n_workers if not args.deepspeed else 1, # To avoid RuntimeError: DataLoader worker exited unexpectedly with exit code 1.
+        num_workers=n_workers,
         persistent_workers=args.persistent_data_loader_workers,
     )
 
@@ -220,30 +223,27 @@ def train(args):
 
     # acceleratorがなんかよろしくやってくれるらしい
     if args.deepspeed:
-        training_models_dict = {}
-        training_models_dict["unet"] = unet
-        if train_text_encoder: training_models_dict["text_encoder"] = text_encoder
-
-        ds_model = train_util.prepare_deepspeed_model(args, **training_models_dict)
-        ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(ds_model, optimizer, train_dataloader, lr_scheduler)
-    
-        training_models = []
-        unet = ds_model.models["unet"]
-        training_models.append(unet)
-        if train_text_encoder:
-            text_encoder = ds_model.models["text_encoder"]
-            training_models.append(text_encoder)
-            
+        if args.train_text_encoder:
+            ds_model = deepspeed_utils.prepare_deepspeed_model(args, unet=unet, text_encoder=text_encoder)
+        else:
+            ds_model = deepspeed_utils.prepare_deepspeed_model(args, unet=unet)
+        ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+            ds_model, optimizer, train_dataloader, lr_scheduler
+        )
+        training_models = [ds_model]
+
     else:
         if train_text_encoder:
             unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
                 unet, text_encoder, optimizer, train_dataloader, lr_scheduler
             )
+            training_models = [unet, text_encoder]
         else:
             unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(unet, optimizer, train_dataloader, lr_scheduler)
+            training_models = [unet]
 
-        if not train_text_encoder:
-            text_encoder.to(accelerator.device, dtype=weight_dtype)  # to avoid 'cpu' vs 'cuda' error
+    if not train_text_encoder:
+        text_encoder.to(accelerator.device, dtype=weight_dtype)  # to avoid 'cpu' vs 'cuda' error
 
     # 実験的機能：勾配も含めたfp16学習を行う　PyTorchにパッチを当ててfp16でのgrad scaleを有効にする
     if args.full_fp16:
@@ -312,8 +312,10 @@ def train(args):
                 if not args.gradient_checkpointing:
                     text_encoder.train(False)
                 text_encoder.requires_grad_(False)
+                if len(training_models) == 2:
+                    training_models = training_models[0]  # remove text_encoder from training_models
 
-            with accelerator.accumulate(unet):
+            with accelerator.accumulate(*training_models):
                 with torch.no_grad():
                     # latentに変換
                     if cache_latents:
@@ -480,6 +482,7 @@ def setup_parser() -> argparse.ArgumentParser:
     train_util.add_sd_models_arguments(parser)
     train_util.add_dataset_arguments(parser, True, False, True)
     train_util.add_training_arguments(parser, True)
+    deepspeed_utils.add_deepspeed_arguments(parser)
     train_util.add_sd_saving_arguments(parser)
     train_util.add_optimizer_arguments(parser)
     config_util.add_config_arguments(parser)
diff --git a/train_network.py b/train_network.py
index af1b7f635..a6ce169a9 100644
--- a/train_network.py
+++ b/train_network.py
@@ -13,13 +13,14 @@
 
 import torch
 from library.device_utils import init_ipex, clean_memory_on_device
+
 init_ipex()
 
 from torch.nn.parallel import DistributedDataParallel as DDP
 
 from accelerate.utils import set_seed
 from diffusers import DDPMScheduler
-from library import model_util
+from library import deepspeed_utils, model_util
 
 import library.train_util as train_util
 from library.train_util import (
@@ -141,6 +142,7 @@ def train(self, args):
         training_started_at = time.time()
         train_util.verify_training_args(args)
         train_util.prepare_dataset_args(args, True)
+        deepspeed_utils.prepare_deepspeed_args(args)
         setup_logging(args, reset=True)
 
         cache_latents = args.cache_latents
@@ -357,7 +359,7 @@ def train(self, args):
             batch_size=1,
             shuffle=True,
             collate_fn=collator,
-            num_workers=n_workers if not args.deepspeed else 1, # To avoid RuntimeError: DataLoader worker exited unexpectedly with exit code 1.
+            num_workers=n_workers,
             persistent_workers=args.persistent_data_loader_workers,
         )
 
@@ -414,22 +416,17 @@ def train(self, args):
 
         # acceleratorがなんかよろしくやってくれるらしい / accelerator will do something good
         if args.deepspeed:
-            training_models_dict = {}
-            if train_unet: training_models_dict["unet"] = unet
-            if train_text_encoder: training_models_dict["text_encoder"] = text_encoders
-            training_models_dict["network"] = network
-
-            ds_model = train_util.prepare_deepspeed_model(args, **training_models_dict)
-            ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(ds_model, optimizer, train_dataloader, lr_scheduler)
-            
-            if train_unet: unet = ds_model.models["unet"]
-            if train_text_encoder:
-                text_encoder = ds_model.models["text_encoder"]
-                if len(ds_model.models["text_encoder"]) > 1:
-                    text_encoders = text_encoder
-                else:
-                    text_encoders = [text_encoder]
-
+            ds_model = deepspeed_utils.prepare_deepspeed_model(
+                args,
+                unet=unet if train_unet else None,
+                text_encoder1=text_encoders[0] if train_text_encoder else None,
+                text_encoder2=text_encoders[1] if train_text_encoder and len(text_encoders) > 1 else None,
+                network=network,
+            )
+            ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+                ds_model, optimizer, train_dataloader, lr_scheduler
+            )
+            training_model = ds_model
         else:
             if train_unet:
                 unet = accelerator.prepare(unet)
@@ -444,7 +441,10 @@ def train(self, args):
             else:
                 pass  # if text_encoder is not trained, no need to prepare. and device and dtype are already set
 
-            network, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(network, optimizer, train_dataloader, lr_scheduler)
+            network, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+                network, optimizer, train_dataloader, lr_scheduler
+            )
+            training_model = network
 
         if args.gradient_checkpointing:
             # according to TI example in Diffusers, train is required
@@ -777,13 +777,13 @@ def remove_model(old_ckpt_name):
 
             for step, batch in enumerate(train_dataloader):
                 current_step.value = global_step
-                with accelerator.accumulate(network):
+                with accelerator.accumulate(training_model):
                     on_step_start(text_encoder, unet)
 
-                    with torch.no_grad():
-                        if "latents" in batch and batch["latents"] is not None:
-                            latents = batch["latents"].to(accelerator.device)
-                        else:
+                    if "latents" in batch and batch["latents"] is not None:
+                        latents = batch["latents"].to(accelerator.device)
+                    else:
+                        with torch.no_grad():
                             # latentに変換
                             latents = vae.encode(batch["images"].to(dtype=vae_dtype)).latent_dist.sample()
 
@@ -791,7 +791,7 @@ def remove_model(old_ckpt_name):
                             if torch.any(torch.isnan(latents)):
                                 accelerator.print("NaN found in latents, replacing with zeros")
                                 latents = torch.nan_to_num(latents, 0, out=latents)
-                        latents = latents * self.vae_scale_factor
+                    latents = latents * self.vae_scale_factor
 
                     # get multiplier for each sample
                     if network_has_multiplier:
@@ -976,6 +976,7 @@ def setup_parser() -> argparse.ArgumentParser:
     train_util.add_sd_models_arguments(parser)
     train_util.add_dataset_arguments(parser, True, True, True)
     train_util.add_training_arguments(parser, True)
+    deepspeed_utils.add_deepspeed_arguments(parser)
     train_util.add_optimizer_arguments(parser)
     config_util.add_config_arguments(parser)
     custom_train_functions.add_custom_train_arguments(parser)

From a9b64ffba8efbb0991a094e38b1f5d5c56680caf Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Tue, 27 Feb 2024 21:43:55 +0900
Subject: [PATCH 17/69] support masked loss in sdxl_train ref #589

---
 README.md     |  4 +++-
 sdxl_train.py | 20 +++++++++++++++++++-
 2 files changed, 22 insertions(+), 2 deletions(-)

diff --git a/README.md b/README.md
index 9cc79cc09..354983c38 100644
--- a/README.md
+++ b/README.md
@@ -251,7 +251,9 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
 
 ### Masked loss
 
-`train_network.py` and `sdxl_train_network.py` now support the masked loss. `--masked_loss` option is added. 
+`train_network.py`, `sdxl_train_network.py` and `sdxl_train.py` now support the masked loss. `--masked_loss` option is added. 
+
+NOTE: `train_network.py` and `sdxl_train.py` are not tested yet.
 
 ControlNet dataset is used to specify the mask. The mask images should be the RGB images. The pixel value 255 in R channel is treated as the mask (the loss is calculated only for the pixels with the mask), and 0 is treated as the non-mask. See details for the dataset specification in the [LLLite documentation](./docs/train_lllite_README.md#preparing-the-dataset).
 
diff --git a/sdxl_train.py b/sdxl_train.py
index e0df263d6..448a160f6 100644
--- a/sdxl_train.py
+++ b/sdxl_train.py
@@ -11,6 +11,7 @@
 
 import torch
 from library.device_utils import init_ipex, clean_memory_on_device
+
 init_ipex()
 
 from accelerate.utils import set_seed
@@ -124,7 +125,7 @@ def train(args):
 
     # データセットを準備する
     if args.dataset_class is None:
-        blueprint_generator = BlueprintGenerator(ConfigSanitizer(True, True, False, True))
+        blueprint_generator = BlueprintGenerator(ConfigSanitizer(True, True, args.masked_loss, True))
         if args.dataset_config is not None:
             logger.info(f"Load dataset config from {args.dataset_config}")
             user_config = config_util.load_user_config(args.dataset_config)
@@ -579,6 +580,16 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
                 ):
                     # do not mean over batch dimension for snr weight or scale v-pred loss
                     loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none")
+
+                    if args.masked_loss:
+                        # mask image is -1 to 1. we need to convert it to 0 to 1
+                        mask_image = batch["conditioning_images"].to(dtype=weight_dtype)[:, 0].unsqueeze(1)  # use R channel
+
+                        # resize to the same size as the loss
+                        mask_image = torch.nn.functional.interpolate(mask_image, size=loss.shape[2:], mode="area")
+                        mask_image = mask_image / 2 + 0.5
+                        loss = loss * mask_image
+
                     loss = loss.mean([1, 2, 3])
 
                     if args.min_snr_gamma:
@@ -780,6 +791,13 @@ def setup_parser() -> argparse.ArgumentParser:
         + f"U-Netの各ブロックの学習率、カンマ区切り、{UNET_NUM_BLOCKS_FOR_BLOCK_LR}個の値",
     )
 
+    # TODO common masked_loss argument
+    parser.add_argument(
+        "--masked_loss",
+        action="store_true",
+        help="apply mask for calculating loss. conditioning_data_dir is required for dataset. / 損失計算時にマスクを適用する。datasetにはconditioning_data_dirが必要",
+    )
+
     return parser
 
 

From 14c9372a38c2d50d8206846fe0aa4406152258de Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sun, 3 Mar 2024 21:47:37 +0900
Subject: [PATCH 18/69] add doc about Colab/rich issue

---
 README.md | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/README.md b/README.md
index e1b6a26c3..7df6fafcc 100644
--- a/README.md
+++ b/README.md
@@ -251,6 +251,7 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
 
 ### Working in progress
 
+- Colab seems to stop with log output. Try specifying `--console_log_simple` option in the training script to disable rich logging.
 - `train_network.py` and `sdxl_train_network.py` are modified to record some dataset settings in the metadata of the trained model (`caption_prefix`, `caption_suffix`, `keep_tokens_separator`, `secondary_separator`, `enable_wildcard`).
 - Some features are added to the dataset subset settings.
   - `secondary_separator` is added to specify the tag separator that is not the target of shuffling or dropping. 
@@ -260,6 +261,8 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
   - The existing features `caption_prefix` and `caption_suffix` can be used together. `caption_prefix` and `caption_suffix` are processed first, and then `enable_wildcard`, `keep_tokens_separator`, shuffling and dropping, and `secondary_separator` are processed in order.
   - The examples are [shown below](#example-of-dataset-settings--データセット設定の記述例).
 
+
+- Colab での動作時、ログ出力で停止してしまうようです。学習スクリプトに `--console_log_simple` オプションを指定し、rich のロギングを無効してお試しください。
 - `train_network.py` および `sdxl_train_network.py` で、学習したモデルのメタデータに一部のデータセット設定が記録されるよう修正しました（`caption_prefix`、`caption_suffix`、`keep_tokens_separator`、`secondary_separator`、`enable_wildcard`）。
 - データセットのサブセット設定にいくつかの機能を追加しました。
   - シャッフルの対象とならないタグ分割識別子の指定 `secondary_separator` を追加しました。`secondary_separator=";;;"` のように指定します。`secondary_separator` で区切ることで、その部分はシャッフル、drop 時にまとめて扱われます。詳しくは記述例をご覧ください。

From 124ec45876a9f07820b42fda0d7ca9019de773d5 Mon Sep 17 00:00:00 2001
From: Horizon1704 <92718180+Horizon1704@users.noreply.github.com>
Date: Sun, 10 Mar 2024 22:53:05 +0800
Subject: [PATCH 19/69] Add "encoding='utf-8'"

---
 library/train_util.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/library/train_util.py b/library/train_util.py
index d2b69edb5..5f23dd13f 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -3474,7 +3474,7 @@ def read_config_from_file(args: argparse.Namespace, parser: argparse.ArgumentPar
         exit(1)
 
     logger.info(f"Loading settings from {config_path}...")
-    with open(config_path, "r") as f:
+    with open(config_path, "r", encoding='utf-8') as f:
         config_dict = toml.load(f)
 
     # combine all sections into one

From 095b8035e63f7c79a232114d8f0e1ec27f431ebc Mon Sep 17 00:00:00 2001
From: gesen2egee <79357052+gesen2egee@users.noreply.github.com>
Date: Sun, 10 Mar 2024 23:33:38 +0800
Subject: [PATCH 20/69] save state on train end

---
 fine_tune.py                     | 2 +-
 library/train_util.py            | 5 +++++
 sdxl_train.py                    | 2 +-
 sdxl_train_control_net_lllite.py | 2 +-
 train_controlnet.py              | 2 +-
 train_db.py                      | 2 +-
 train_network.py                 | 2 +-
 train_textual_inversion.py       | 2 +-
 train_textual_inversion_XTI.py   | 2 +-
 9 files changed, 13 insertions(+), 8 deletions(-)

diff --git a/fine_tune.py b/fine_tune.py
index 875a91951..46f128287 100644
--- a/fine_tune.py
+++ b/fine_tune.py
@@ -457,7 +457,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
 
     accelerator.end_training()
 
-    if args.save_state and is_main_process:
+    if is_main_process and (args.save_state or args.save_state_on_train_end):        
         train_util.save_state_on_train_end(args, accelerator)
 
     del accelerator  # この後メモリを使うのでこれは消す
diff --git a/library/train_util.py b/library/train_util.py
index d2b69edb5..b3ca15f55 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -2890,6 +2890,11 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
         action="store_true",
         help="save training state additionally (including optimizer states etc.) / optimizerなど学習状態も含めたstateを追加で保存する",
     )
+    parser.add_argument(
+        "--save_state_on_train_end",
+        action="store_true",
+        help="save training state additionally (including optimizer states etc.) on train end / optimizerなど学習状態も含めたstateを追加で保存する",
+    )   
     parser.add_argument("--resume", type=str, default=None, help="saved state to resume training / 学習再開するモデルのstate")
 
     parser.add_argument("--train_batch_size", type=int, default=1, help="batch size for training / 学習時のバッチサイズ")
diff --git a/sdxl_train.py b/sdxl_train.py
index e0df263d6..107bb9451 100644
--- a/sdxl_train.py
+++ b/sdxl_train.py
@@ -712,7 +712,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
 
     accelerator.end_training()
 
-    if args.save_state:  # and is_main_process:
+    if args.save_state or args.save_state_on_train_end:        
         train_util.save_state_on_train_end(args, accelerator)
 
     del accelerator  # この後メモリを使うのでこれは消す
diff --git a/sdxl_train_control_net_lllite.py b/sdxl_train_control_net_lllite.py
index 1e5f92349..e99b4e35c 100644
--- a/sdxl_train_control_net_lllite.py
+++ b/sdxl_train_control_net_lllite.py
@@ -549,7 +549,7 @@ def remove_model(old_ckpt_name):
 
     accelerator.end_training()
 
-    if is_main_process and args.save_state:
+    if is_main_process and (args.save_state or args.save_state_on_train_end):
         train_util.save_state_on_train_end(args, accelerator)
 
     if is_main_process:
diff --git a/train_controlnet.py b/train_controlnet.py
index dc73a91c8..e44f08853 100644
--- a/train_controlnet.py
+++ b/train_controlnet.py
@@ -565,7 +565,7 @@ def remove_model(old_ckpt_name):
 
     accelerator.end_training()
 
-    if is_main_process and args.save_state:
+    if is_main_process and (args.save_state or args.save_state_on_train_end):
         train_util.save_state_on_train_end(args, accelerator)
 
     # del accelerator  # この後メモリを使うのでこれは消す→printで使うので消さずにおく
diff --git a/train_db.py b/train_db.py
index 8d36097a5..41a9a7b99 100644
--- a/train_db.py
+++ b/train_db.py
@@ -444,7 +444,7 @@ def train(args):
 
     accelerator.end_training()
 
-    if args.save_state and is_main_process:
+    if is_main_process and (args.save_state or args.save_state_on_train_end):
         train_util.save_state_on_train_end(args, accelerator)
 
     del accelerator  # この後メモリを使うのでこれは消す
diff --git a/train_network.py b/train_network.py
index e0fa69458..4707d5ae5 100644
--- a/train_network.py
+++ b/train_network.py
@@ -935,7 +935,7 @@ def remove_model(old_ckpt_name):
 
         accelerator.end_training()
 
-        if is_main_process and args.save_state:
+        if is_main_process and args.save_state or args.save_state_on_train_end:
             train_util.save_state_on_train_end(args, accelerator)
 
         if is_main_process:
diff --git a/train_textual_inversion.py b/train_textual_inversion.py
index df1d8485a..0266bc143 100644
--- a/train_textual_inversion.py
+++ b/train_textual_inversion.py
@@ -732,7 +732,7 @@ def remove_model(old_ckpt_name):
 
         accelerator.end_training()
 
-        if args.save_state and is_main_process:
+        if is_main_process and (args.save_state or args.save_state_on_train_end):
             train_util.save_state_on_train_end(args, accelerator)
 
         if is_main_process:
diff --git a/train_textual_inversion_XTI.py b/train_textual_inversion_XTI.py
index 695fad2a8..ad7c267eb 100644
--- a/train_textual_inversion_XTI.py
+++ b/train_textual_inversion_XTI.py
@@ -586,7 +586,7 @@ def remove_model(old_ckpt_name):
 
     accelerator.end_training()
 
-    if args.save_state and is_main_process:
+    if is_main_process and (args.save_state or args.save_state_on_train_end):
         train_util.save_state_on_train_end(args, accelerator)
 
     updated_embs = text_encoder.get_input_embeddings().weight[token_ids_XTI].data.detach().clone()

From d282c450026dcfd5f1fd5856f5087ebaed47be46 Mon Sep 17 00:00:00 2001
From: gesen2egee <79357052+gesen2egee@users.noreply.github.com>
Date: Mon, 11 Mar 2024 23:56:09 +0800
Subject: [PATCH 21/69] Update train_network.py

---
 train_network.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/train_network.py b/train_network.py
index 4707d5ae5..3db583f1d 100644
--- a/train_network.py
+++ b/train_network.py
@@ -935,7 +935,7 @@ def remove_model(old_ckpt_name):
 
         accelerator.end_training()
 
-        if is_main_process and args.save_state or args.save_state_on_train_end:
+        if is_main_process and (args.save_state or args.save_state_on_train_end):
             train_util.save_state_on_train_end(args, accelerator)
 
         if is_main_process:

From 948029fe61d9142f88374d6701223bf9f7ee5d47 Mon Sep 17 00:00:00 2001
From: kblueleaf <apolloyeh0123@gmail.com>
Date: Tue, 12 Mar 2024 19:11:45 +0800
Subject: [PATCH 22/69] random ip_noise_gamma strength

---
 library/train_util.py | 13 ++++++++++++-
 1 file changed, 12 insertions(+), 1 deletion(-)

diff --git a/library/train_util.py b/library/train_util.py
index b71e4edc6..aa2d9b90b 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -3100,6 +3100,13 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
         help="enable input perturbation noise. used for regularization. recommended value: around 0.1 (from arxiv.org/abs/2301.11706) "
         + "/  input perturbation noiseを有効にする。正則化に使用される。推奨値: 0.1程度 (arxiv.org/abs/2301.11706 より)",
     )
+    parser.add_argument(
+        "--ip_noise_gamma_random_strength",
+        type=bool,
+        default=False,
+        help="Use random strength between 0~ip_noise_gamma for input perturbation noise."
+        + "/ input perturbation noiseにおいて、0からip_noise_gammaの間でランダムな強度を使用します。",
+    )
     # parser.add_argument(
     #     "--perlin_noise",
     #     type=int,
@@ -4673,7 +4680,11 @@ def get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents):
     # Add noise to the latents according to the noise magnitude at each timestep
     # (this is the forward diffusion process)
     if args.ip_noise_gamma:
-        noisy_latents = noise_scheduler.add_noise(latents, noise + args.ip_noise_gamma * torch.randn_like(latents), timesteps)
+        if args.ip_noise_gamma_random_strength:
+            strength = torch.rand(1, device=latents.device) * args.ip_noise_gamma
+        else:
+            strength = args.ip_noise_gamma
+        noisy_latents = noise_scheduler.add_noise(latents, noise + strength * torch.randn_like(latents), timesteps)
     else:
         noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
 

From 86399407b2bb5a93d691846acfa88e7ba38ae70d Mon Sep 17 00:00:00 2001
From: kblueleaf <apolloyeh0123@gmail.com>
Date: Tue, 12 Mar 2024 19:14:01 +0800
Subject: [PATCH 23/69] random noise_offset strength

---
 library/train_util.py | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

diff --git a/library/train_util.py b/library/train_util.py
index aa2d9b90b..5282b5240 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -3087,6 +3087,12 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
         default=None,
         help="enable noise offset with this value (if enabled, around 0.1 is recommended) / Noise offsetを有効にしてこの値を設定する（有効にする場合は0.1程度を推奨）",
     )
+    parser.add_argument(
+        "--noise_offset_random_strength",
+        type=bool,
+        default=False,
+        help="use random strength between 0~noise_offset for noise offset. / noise offsetにおいて、0からnoise_offsetの間でランダムな強度を使用します。",
+    )
     parser.add_argument(
         "--multires_noise_iterations",
         type=int,
@@ -4663,7 +4669,11 @@ def get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents):
     # Sample noise that we'll add to the latents
     noise = torch.randn_like(latents, device=latents.device)
     if args.noise_offset:
-        noise = custom_train_functions.apply_noise_offset(latents, noise, args.noise_offset, args.adaptive_noise_scale)
+        if args.noise_offset_random_strength:
+            noise_offset = torch.rand(1, device=latents.device) * args.noise_offset
+        else:
+            noise_offset = args.noise_offset
+        noise = custom_train_functions.apply_noise_offset(latents, noise, noise_offset, args.adaptive_noise_scale)
     if args.multires_noise_iterations:
         noise = custom_train_functions.pyramid_noise_like(
             noise, latents.device, args.multires_noise_iterations, args.multires_noise_discount

From 53954a1e2e05648bae6eb479720402968029cd3d Mon Sep 17 00:00:00 2001
From: kblueleaf <apolloyeh0123@gmail.com>
Date: Tue, 12 Mar 2024 19:24:27 +0800
Subject: [PATCH 24/69] use correct settings for parser

---
 library/train_util.py | 6 ++----
 1 file changed, 2 insertions(+), 4 deletions(-)

diff --git a/library/train_util.py b/library/train_util.py
index 5282b5240..73a768672 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -3089,8 +3089,7 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
     )
     parser.add_argument(
         "--noise_offset_random_strength",
-        type=bool,
-        default=False,
+        action="store_true",
         help="use random strength between 0~noise_offset for noise offset. / noise offsetにおいて、0からnoise_offsetの間でランダムな強度を使用します。",
     )
     parser.add_argument(
@@ -3108,8 +3107,7 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
     )
     parser.add_argument(
         "--ip_noise_gamma_random_strength",
-        type=bool,
-        default=False,
+        action="store_true",
         help="Use random strength between 0~ip_noise_gamma for input perturbation noise."
         + "/ input perturbation noiseにおいて、0からip_noise_gammaの間でランダムな強度を使用します。",
     )

From 443f02942cfefd6e1899849f563580508d118ce0 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Fri, 15 Mar 2024 21:35:14 +0900
Subject: [PATCH 25/69] fix doc

---
 README.md | 106 ------------------------------------------------------
 1 file changed, 106 deletions(-)

diff --git a/README.md b/README.md
index cc5aca505..927d7a42e 100644
--- a/README.md
+++ b/README.md
@@ -355,112 +355,6 @@ It becomes `1girl, hatsune miku, vocaloid, microphone, stage, white shirt, best
 `1girl, hatsune miku, vocaloid, microphone, stage, white shirt, best quality, rating: general` や `1girl, hatsune miku, vocaloid, white shirt, smile, stage, microphone, best quality, rating: general` などになります。
 
 
-### Working in progress
-
-- Colab seems to stop with log output. Try specifying `--console_log_simple` option in the training script to disable rich logging.
-- `train_network.py` and `sdxl_train_network.py` are modified to record some dataset settings in the metadata of the trained model (`caption_prefix`, `caption_suffix`, `keep_tokens_separator`, `secondary_separator`, `enable_wildcard`).
-- Some features are added to the dataset subset settings.
-  - `secondary_separator` is added to specify the tag separator that is not the target of shuffling or dropping. 
-    - Specify `secondary_separator=";;;"`. When you specify `secondary_separator`, the part is not shuffled or dropped. See the example below.
-  - `enable_wildcard` is added. When set to `true`, the wildcard notation `{aaa|bbb|ccc}` can be used. See the example below.
-  - `keep_tokens_separator` is updated to be used twice in the caption. When you specify `keep_tokens_separator="|||"`, the part divided by the second `|||` is not shuffled or dropped and remains at the end.
-  - The existing features `caption_prefix` and `caption_suffix` can be used together. `caption_prefix` and `caption_suffix` are processed first, and then `enable_wildcard`, `keep_tokens_separator`, shuffling and dropping, and `secondary_separator` are processed in order.
-  - The examples are [shown below](#example-of-dataset-settings--データセット設定の記述例).
-
-
-- Colab での動作時、ログ出力で停止してしまうようです。学習スクリプトに `--console_log_simple` オプションを指定し、rich のロギングを無効してお試しください。
-- `train_network.py` および `sdxl_train_network.py` で、学習したモデルのメタデータに一部のデータセット設定が記録されるよう修正しました（`caption_prefix`、`caption_suffix`、`keep_tokens_separator`、`secondary_separator`、`enable_wildcard`）。
-- データセットのサブセット設定にいくつかの機能を追加しました。
-  - シャッフルの対象とならないタグ分割識別子の指定 `secondary_separator` を追加しました。`secondary_separator=";;;"` のように指定します。`secondary_separator` で区切ることで、その部分はシャッフル、drop 時にまとめて扱われます。詳しくは記述例をご覧ください。
-  - `enable_wildcard` を追加しました。`true` にするとワイルドカード記法 `{aaa|bbb|ccc}` が使えます。詳しくは記述例をご覧ください。
-  - `keep_tokens_separator` をキャプション内に 2 つ使えるようにしました。たとえば `keep_tokens_separator="|||"` と指定したとき、`1girl, hatsune miku, vocaloid ||| stage, mic ||| best quality, rating: general` とキャプションを指定すると、二番目の `|||` で分割された部分はシャッフル、drop されず末尾に残ります。
-  - 既存の機能 `caption_prefix` と `caption_suffix` とあわせて使えます。`caption_prefix` と `caption_suffix` は一番最初に処理され、その後、ワイルドカード、`keep_tokens_separator`、シャッフルおよび drop、`secondary_separator` の順に処理されます。
-
-#### Example of dataset settings / データセット設定の記述例:
-
-```toml
-[general]
-flip_aug = true
-color_aug = false
-resolution = [1024, 1024]
-
-[[datasets]]
-batch_size = 6
-enable_bucket = true
-bucket_no_upscale = true
-caption_extension = ".txt"
-keep_tokens_separator= "|||"
-shuffle_caption = true
-caption_tag_dropout_rate = 0.1
-secondary_separator = ";;;" # subset 側に書くこともできます / can be written in the subset side
-enable_wildcard = true # 同上 / same as above
-
-  [[datasets.subsets]]
-  image_dir = "/path/to/image_dir"
-  num_repeats = 1
-
-  # ||| の前後はカンマは不要です（自動的に追加されます） / No comma is required before and after ||| (it is added automatically)
-  caption_prefix = "1girl, hatsune miku, vocaloid |||" 
-  
-  # ||| の後はシャッフル、drop されず残ります / After |||, it is not shuffled or dropped and remains
-  # 単純に文字列として連結されるので、カンマなどは自分で入れる必要があります / It is simply concatenated as a string, so you need to put commas yourself
-  caption_suffix = ", anime screencap ||| masterpiece, rating: general"
-```
-
-#### Example of caption, secondary_separator notation: `secondary_separator = ";;;"`
-
-```txt
-1girl, hatsune miku, vocaloid, upper body, looking at viewer, sky;;;cloud;;;day, outdoors
-```
-The part `sky;;;cloud;;;day` is replaced with `sky,cloud,day` without shuffling or dropping. When shuffling and dropping are enabled, it is processed as a whole (as one tag). For example, it becomes `vocaloid, 1girl, upper body, sky,cloud,day, outdoors, hatsune miku` (shuffled) or `vocaloid, 1girl, outdoors, looking at viewer, upper body, hatsune miku` (dropped).
-
-#### Example of caption, enable_wildcard notation: `enable_wildcard = true`
-
-```txt
-1girl, hatsune miku, vocaloid, upper body, looking at viewer, {simple|white} background
-```
-`simple` or `white` is randomly selected, and it becomes `simple background` or `white background`.
-
-```txt
-1girl, hatsune miku, vocaloid, {{retro style}}
-```
-If you want to include `{` or `}` in the tag string, double them like `{{` or `}}` (in this example, the actual caption used for training is `{retro style}`).
-
-#### Example of caption, `keep_tokens_separator` notation: `keep_tokens_separator = "|||"`
-
-```txt
-1girl, hatsune miku, vocaloid ||| stage, microphone, white shirt, smile ||| best quality, rating: general
-```
-It becomes `1girl, hatsune miku, vocaloid, microphone, stage, white shirt, best quality, rating: general` or `1girl, hatsune miku, vocaloid, white shirt, smile, stage, microphone, best quality, rating: general` etc.
-
-
-#### キャプション記述例、secondary_separator 記法：`secondary_separator = ";;;"` の場合
-
-```txt
-1girl, hatsune miku, vocaloid, upper body, looking at viewer, sky;;;cloud;;;day, outdoors
-```
-`sky;;;cloud;;;day` の部分はシャッフル、drop されず `sky,cloud,day` に置換されます。シャッフル、drop が有効な場合、まとめて（一つのタグとして）処理されます。つまり `vocaloid, 1girl, upper body, sky,cloud,day, outdoors, hatsune miku` （シャッフル）や `vocaloid, 1girl, outdoors, looking at viewer, upper body, hatsune miku` （drop されたケース）などになります。
-
-#### キャプション記述例、ワイルドカード記法： `enable_wildcard = true` の場合
-
-```txt
-1girl, hatsune miku, vocaloid, upper body, looking at viewer, {simple|white} background
-```
-ランダムに `simple` または `white` が選ばれ、`simple background` または `white background` になります。
-
-```txt
-1girl, hatsune miku, vocaloid, {{retro style}}
-```
-タグ文字列に `{` や `}` そのものを含めたい場合は `{{` や `}}` のように二つ重ねてください（この例では実際に学習に用いられるキャプションは `{retro style}` になります）。
-
-#### キャプション記述例、`keep_tokens_separator` 記法： `keep_tokens_separator = "|||"` の場合
-
-```txt
-1girl, hatsune miku, vocaloid ||| stage, microphone, white shirt, smile ||| best quality, rating: general
-```
-`1girl, hatsune miku, vocaloid, microphone, stage, white shirt, best quality, rating: general` や `1girl, hatsune miku, vocaloid, white shirt, smile, stage, microphone, best quality, rating: general` などになります。
-
-
 ### Mar 15, 2024 / 2024/3/15: v0.8.5
 
 - Fixed a bug that the value of timestep embedding during SDXL training was incorrect.

From 7081a0cf0f1ca1a543edf7cab10c4c7d497348ca Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sun, 17 Mar 2024 18:09:15 +0900
Subject: [PATCH 26/69] extension of src image could be different than target
 image

---
 library/train_util.py | 26 ++++++++++++++++----------
 1 file changed, 16 insertions(+), 10 deletions(-)

diff --git a/library/train_util.py b/library/train_util.py
index 7fe5bc56e..0f8cf9eea 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -1863,7 +1863,7 @@ def __init__(
 
         # assert all conditioning data exists
         missing_imgs = []
-        cond_imgs_with_img = set()
+        cond_imgs_with_pair = set()
         for image_key, info in self.dreambooth_dataset_delegate.image_data.items():
             db_subset = self.dreambooth_dataset_delegate.image_to_subset[image_key]
             subset = None
@@ -1877,23 +1877,29 @@ def __init__(
                 logger.warning(f"not directory: {subset.conditioning_data_dir}")
                 continue
 
-            img_basename = os.path.basename(info.absolute_path)
-            ctrl_img_path = os.path.join(subset.conditioning_data_dir, img_basename)
-            if not os.path.exists(ctrl_img_path):
+            img_basename = os.path.splitext(os.path.basename(info.absolute_path))[0]
+            ctrl_img_path = glob_images(subset.conditioning_data_dir, img_basename)
+            if len(ctrl_img_path) < 1:
                 missing_imgs.append(img_basename)
+                continue
+            ctrl_img_path = ctrl_img_path[0]
+            ctrl_img_path = os.path.abspath(ctrl_img_path)  # normalize path
 
             info.cond_img_path = ctrl_img_path
-            cond_imgs_with_img.add(ctrl_img_path)
+            cond_imgs_with_pair.add(os.path.splitext(ctrl_img_path)[0])  # remove extension because Windows is case insensitive
 
         extra_imgs = []
         for subset in subsets:
             conditioning_img_paths = glob_images(subset.conditioning_data_dir, "*")
-            extra_imgs.extend(
-                [cond_img_path for cond_img_path in conditioning_img_paths if cond_img_path not in cond_imgs_with_img]
-            )
+            conditioning_img_paths = [os.path.abspath(p) for p in conditioning_img_paths]  # normalize path
+            extra_imgs.extend([p for p in conditioning_img_paths if os.path.splitext(p)[0] not in cond_imgs_with_pair])
 
-        assert len(missing_imgs) == 0, f"missing conditioning data for {len(missing_imgs)} images: {missing_imgs}"
-        assert len(extra_imgs) == 0, f"extra conditioning data for {len(extra_imgs)} images: {extra_imgs}"
+        assert (
+            len(missing_imgs) == 0
+        ), f"missing conditioning data for {len(missing_imgs)} images / 制御用画像が見つかりませんでした: {missing_imgs}"
+        assert (
+            len(extra_imgs) == 0
+        ), f"extra conditioning data for {len(extra_imgs)} images / 余分な制御用画像があります: {extra_imgs}"
 
         self.conditioning_image_transforms = IMAGE_TRANSFORMS
 

From 3419c3de0d0ff8cba1d74444ece23608614f3c5b Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sun, 17 Mar 2024 19:30:20 +0900
Subject: [PATCH 27/69] common masked loss func, apply to all training script

---
 docs/train_lllite_README-ja.md    |  8 ++++++--
 docs/train_lllite_README.md       |  4 +++-
 library/config_util.py            |  5 ++++-
 library/custom_train_functions.py | 24 ++++++++++++++++++++----
 library/train_util.py             | 16 ++++++++++++++++
 sdxl_train.py                     | 21 ++++-----------------
 train_db.py                       |  7 ++++++-
 train_network.py                  | 17 +++--------------
 train_textual_inversion.py        |  7 ++++++-
 train_textual_inversion_XTI.py    |  7 ++++++-
 10 files changed, 74 insertions(+), 42 deletions(-)

diff --git a/docs/train_lllite_README-ja.md b/docs/train_lllite_README-ja.md
index dbdc1fea2..1f6a78d5c 100644
--- a/docs/train_lllite_README-ja.md
+++ b/docs/train_lllite_README-ja.md
@@ -21,9 +21,13 @@ ComfyUIのカスタムノードを用意しています。: https://github.com/k
 ## モデルの学習
 
 ### データセットの準備
-通常のdatasetに加え、`conditioning_data_dir` で指定したディレクトリにconditioning imageを格納してください。conditioning imageは学習用画像と同じbasenameを持つ必要があります。また、conditioning imageは学習用画像と同じサイズに自動的にリサイズされます。conditioning imageにはキャプションファイルは不要です。
+DreamBooth 方式の dataset で、`conditioning_data_dir` で指定したディレクトリにconditioning imageを格納してください。
 
-たとえば DreamBooth 方式でキャプションファイルを用いる場合の設定ファイルは以下のようになります。
+（finetuning 方式の dataset はサポートしていません。）
+
+conditioning imageは学習用画像と同じbasenameを持つ必要があります。また、conditioning imageは学習用画像と同じサイズに自動的にリサイズされます。conditioning imageにはキャプションファイルは不要です。
+
+たとえば、キャプションにフォルダ名ではなくキャプションファイルを用いる場合の設定ファイルは以下のようになります。
 
 ```toml
 [[datasets.subsets]]
diff --git a/docs/train_lllite_README.md b/docs/train_lllite_README.md
index 04dc12da2..a05f87f5f 100644
--- a/docs/train_lllite_README.md
+++ b/docs/train_lllite_README.md
@@ -26,7 +26,9 @@ Due to the limitations of the inference environment, only CrossAttention (attn1
 
 ### Preparing the dataset
 
-In addition to the normal dataset, please store the conditioning image in the directory specified by `conditioning_data_dir`. The conditioning image must have the same basename as the training image. The conditioning image will be automatically resized to the same size as the training image. The conditioning image does not require a caption file.
+In addition to the normal DreamBooth method dataset, please store the conditioning image in the directory specified by `conditioning_data_dir`. The conditioning image must have the same basename as the training image. The conditioning image will be automatically resized to the same size as the training image. The conditioning image does not require a caption file.
+
+(We do not support the finetuning method dataset.)
 
 ```toml
 [[datasets.subsets]]
diff --git a/library/config_util.py b/library/config_util.py
index edc6a5385..26daeb472 100644
--- a/library/config_util.py
+++ b/library/config_util.py
@@ -323,7 +323,10 @@ def validate_flex_dataset(dataset_config: dict):
 
             self.dataset_schema = validate_flex_dataset
         elif support_dreambooth:
-            self.dataset_schema = self.db_dataset_schema
+            if support_controlnet:
+                self.dataset_schema = self.cn_dataset_schema
+            else:
+                self.dataset_schema = self.db_dataset_schema
         elif support_finetuning:
             self.dataset_schema = self.ft_dataset_schema
         elif support_controlnet:
diff --git a/library/custom_train_functions.py b/library/custom_train_functions.py
index a56474622..406e0e36e 100644
--- a/library/custom_train_functions.py
+++ b/library/custom_train_functions.py
@@ -3,11 +3,14 @@
 import random
 import re
 from typing import List, Optional, Union
-from .utils import setup_logging
+from .utils import setup_logging
+
 setup_logging()
-import logging
+import logging
+
 logger = logging.getLogger(__name__)
 
+
 def prepare_scheduler_for_custom_training(noise_scheduler, device):
     if hasattr(noise_scheduler, "all_snr"):
         return
@@ -64,7 +67,7 @@ def apply_snr_weight(loss, timesteps, noise_scheduler, gamma, v_prediction=False
     snr = torch.stack([noise_scheduler.all_snr[t] for t in timesteps])
     min_snr_gamma = torch.minimum(snr, torch.full_like(snr, gamma))
     if v_prediction:
-        snr_weight = torch.div(min_snr_gamma, snr+1).float().to(loss.device)
+        snr_weight = torch.div(min_snr_gamma, snr + 1).float().to(loss.device)
     else:
         snr_weight = torch.div(min_snr_gamma, snr).float().to(loss.device)
     loss = loss * snr_weight
@@ -92,13 +95,15 @@ def add_v_prediction_like_loss(loss, timesteps, noise_scheduler, v_pred_like_los
     loss = loss + loss / scale * v_pred_like_loss
     return loss
 
+
 def apply_debiased_estimation(loss, timesteps, noise_scheduler):
     snr_t = torch.stack([noise_scheduler.all_snr[t] for t in timesteps])  # batch_size
     snr_t = torch.minimum(snr_t, torch.ones_like(snr_t) * 1000)  # if timestep is 0, snr_t is inf, so limit it to 1000
-    weight = 1/torch.sqrt(snr_t)
+    weight = 1 / torch.sqrt(snr_t)
     loss = weight * loss
     return loss
 
+
 # TODO train_utilと分散しているのでどちらかに寄せる
 
 
@@ -474,6 +479,17 @@ def apply_noise_offset(latents, noise, noise_offset, adaptive_noise_scale):
     return noise
 
 
+def apply_masked_loss(loss, batch):
+    # mask image is -1 to 1. we need to convert it to 0 to 1
+    mask_image = batch["conditioning_images"].to(dtype=loss.dtype)[:, 0].unsqueeze(1)  # use R channel
+
+    # resize to the same size as the loss
+    mask_image = torch.nn.functional.interpolate(mask_image, size=loss.shape[2:], mode="area")
+    mask_image = mask_image / 2 + 0.5
+    loss = loss * mask_image
+    return loss
+
+
 """
 ##########################################
 # Perlin Noise
diff --git a/library/train_util.py b/library/train_util.py
index 0f8cf9eea..1d9f8bf82 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -3028,6 +3028,7 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
         "--full_bf16", action="store_true", help="bf16 training including gradients / 勾配も含めてbf16で学習する"
     )  # TODO move to SDXL training, because it is not supported by SD1/2
     parser.add_argument("--fp8_base", action="store_true", help="use fp8 for base model / base modelにfp8を使う")
+
     parser.add_argument(
         "--ddp_timeout",
         type=int,
@@ -3090,6 +3091,7 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
         default=None,
         help="specify WandB API key to log in before starting training (optional). / WandB APIキーを指定して学習開始前にログインする（オプション）",
     )
+
     parser.add_argument(
         "--noise_offset",
         type=float,
@@ -3252,6 +3254,20 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
         )
 
 
+def add_masked_loss_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--conditioning_data_dir",
+        type=str,
+        default=None,
+        help="conditioning data directory / 条件付けデータのディレクトリ",
+    )
+    parser.add_argument(
+        "--masked_loss",
+        action="store_true",
+        help="apply mask for calculating loss. conditioning_data_dir is required for dataset. / 損失計算時にマスクを適用する。datasetにはconditioning_data_dirが必要",
+    )
+
+
 def verify_training_args(args: argparse.Namespace):
     r"""
     Verify training arguments. Also reflect highvram option to global variable
diff --git a/sdxl_train.py b/sdxl_train.py
index 448a160f6..f8aa46081 100644
--- a/sdxl_train.py
+++ b/sdxl_train.py
@@ -40,6 +40,7 @@
     scale_v_prediction_loss_like_noise_prediction,
     add_v_prediction_like_loss,
     apply_debiased_estimation,
+    apply_masked_loss,
 )
 from library.sdxl_original_unet import SdxlUNet2DConditionModel
 
@@ -577,19 +578,12 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
                     or args.scale_v_pred_loss_like_noise_pred
                     or args.v_pred_like_loss
                     or args.debiased_estimation_loss
+                    or args.masked_loss
                 ):
                     # do not mean over batch dimension for snr weight or scale v-pred loss
                     loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none")
-
                     if args.masked_loss:
-                        # mask image is -1 to 1. we need to convert it to 0 to 1
-                        mask_image = batch["conditioning_images"].to(dtype=weight_dtype)[:, 0].unsqueeze(1)  # use R channel
-
-                        # resize to the same size as the loss
-                        mask_image = torch.nn.functional.interpolate(mask_image, size=loss.shape[2:], mode="area")
-                        mask_image = mask_image / 2 + 0.5
-                        loss = loss * mask_image
-
+                        loss = apply_masked_loss(loss, batch)
                     loss = loss.mean([1, 2, 3])
 
                     if args.min_snr_gamma:
@@ -755,6 +749,7 @@ def setup_parser() -> argparse.ArgumentParser:
     train_util.add_sd_models_arguments(parser)
     train_util.add_dataset_arguments(parser, True, True, True)
     train_util.add_training_arguments(parser, False)
+    train_util.add_masked_loss_arguments(parser)
     train_util.add_sd_saving_arguments(parser)
     train_util.add_optimizer_arguments(parser)
     config_util.add_config_arguments(parser)
@@ -790,14 +785,6 @@ def setup_parser() -> argparse.ArgumentParser:
         help=f"learning rates for each block of U-Net, comma-separated, {UNET_NUM_BLOCKS_FOR_BLOCK_LR} values / "
         + f"U-Netの各ブロックの学習率、カンマ区切り、{UNET_NUM_BLOCKS_FOR_BLOCK_LR}個の値",
     )
-
-    # TODO common masked_loss argument
-    parser.add_argument(
-        "--masked_loss",
-        action="store_true",
-        help="apply mask for calculating loss. conditioning_data_dir is required for dataset. / 損失計算時にマスクを適用する。datasetにはconditioning_data_dirが必要",
-    )
-
     return parser
 
 
diff --git a/train_db.py b/train_db.py
index 8d36097a5..213df1516 100644
--- a/train_db.py
+++ b/train_db.py
@@ -12,6 +12,7 @@
 
 import torch
 from library.device_utils import init_ipex, clean_memory_on_device
+
 init_ipex()
 
 from accelerate.utils import set_seed
@@ -32,6 +33,7 @@
     apply_noise_offset,
     scale_v_prediction_loss_like_noise_prediction,
     apply_debiased_estimation,
+    apply_masked_loss,
 )
 from library.utils import setup_logging, add_logging_arguments
 
@@ -57,7 +59,7 @@ def train(args):
 
     # データセットを準備する
     if args.dataset_class is None:
-        blueprint_generator = BlueprintGenerator(ConfigSanitizer(True, False, False, True))
+        blueprint_generator = BlueprintGenerator(ConfigSanitizer(True, False, args.masked_loss, True))
         if args.dataset_config is not None:
             logger.info(f"Load dataset config from {args.dataset_config}")
             user_config = config_util.load_user_config(args.dataset_config)
@@ -339,6 +341,8 @@ def train(args):
                     target = noise
 
                 loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none")
+                if args.masked_loss:
+                    loss = apply_masked_loss(loss, batch)
                 loss = loss.mean([1, 2, 3])
 
                 loss_weights = batch["loss_weights"]  # 各sampleごとのweight
@@ -464,6 +468,7 @@ def setup_parser() -> argparse.ArgumentParser:
     train_util.add_sd_models_arguments(parser)
     train_util.add_dataset_arguments(parser, True, False, True)
     train_util.add_training_arguments(parser, True)
+    train_util.add_masked_loss_arguments(parser)
     train_util.add_sd_saving_arguments(parser)
     train_util.add_optimizer_arguments(parser)
     config_util.add_config_arguments(parser)
diff --git a/train_network.py b/train_network.py
index f5617986c..05522070b 100644
--- a/train_network.py
+++ b/train_network.py
@@ -40,6 +40,7 @@
     scale_v_prediction_loss_like_noise_prediction,
     add_v_prediction_like_loss,
     apply_debiased_estimation,
+    apply_masked_loss,
 )
 from library.utils import setup_logging, add_logging_arguments
 
@@ -835,16 +836,8 @@ def remove_model(old_ckpt_name):
                         target = noise
 
                     loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none")
-
                     if args.masked_loss:
-                        # mask image is -1 to 1. we need to convert it to 0 to 1
-                        mask_image = batch["conditioning_images"].to(dtype=weight_dtype)[:, 0].unsqueeze(1)  # use R channel
-
-                        # resize to the same size as the loss
-                        mask_image = torch.nn.functional.interpolate(mask_image, size=loss.shape[2:], mode="area")
-                        mask_image = mask_image / 2 + 0.5
-                        loss = loss * mask_image
-
+                        loss = apply_masked_loss(loss, batch)
                     loss = loss.mean([1, 2, 3])
 
                     loss_weights = batch["loss_weights"]  # 各sampleごとのweight
@@ -968,6 +961,7 @@ def setup_parser() -> argparse.ArgumentParser:
     train_util.add_sd_models_arguments(parser)
     train_util.add_dataset_arguments(parser, True, True, True)
     train_util.add_training_arguments(parser, True)
+    train_util.add_masked_loss_arguments(parser)
     train_util.add_optimizer_arguments(parser)
     config_util.add_config_arguments(parser)
     custom_train_functions.add_custom_train_arguments(parser)
@@ -1061,11 +1055,6 @@ def setup_parser() -> argparse.ArgumentParser:
         action="store_true",
         help="do not use fp16/bf16 VAE in mixed precision (use float VAE) / mixed precisionでも fp16/bf16 VAEを使わずfloat VAEを使う",
     )
-    parser.add_argument(
-        "--masked_loss",
-        action="store_true",
-        help="apply mask for calculating loss. conditioning_data_dir is required for dataset. / 損失計算時にマスクを適用する。datasetにはconditioning_data_dirが必要",
-    )
     return parser
 
 
diff --git a/train_textual_inversion.py b/train_textual_inversion.py
index df1d8485a..7697b9672 100644
--- a/train_textual_inversion.py
+++ b/train_textual_inversion.py
@@ -8,6 +8,7 @@
 
 import torch
 from library.device_utils import init_ipex, clean_memory_on_device
+
 init_ipex()
 
 from accelerate.utils import set_seed
@@ -29,6 +30,7 @@
     scale_v_prediction_loss_like_noise_prediction,
     add_v_prediction_like_loss,
     apply_debiased_estimation,
+    apply_masked_loss,
 )
 from library.utils import setup_logging, add_logging_arguments
 
@@ -268,7 +270,7 @@ def train(self, args):
 
         # データセットを準備する
         if args.dataset_class is None:
-            blueprint_generator = BlueprintGenerator(ConfigSanitizer(True, True, False, False))
+            blueprint_generator = BlueprintGenerator(ConfigSanitizer(True, True, args.masked_loss, False))
             if args.dataset_config is not None:
                 accelerator.print(f"Load dataset config from {args.dataset_config}")
                 user_config = config_util.load_user_config(args.dataset_config)
@@ -586,6 +588,8 @@ def remove_model(old_ckpt_name):
                         target = noise
 
                     loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none")
+                    if args.masked_loss:
+                        loss = apply_masked_loss(loss, batch)
                     loss = loss.mean([1, 2, 3])
 
                     loss_weights = batch["loss_weights"]  # 各sampleごとのweight
@@ -749,6 +753,7 @@ def setup_parser() -> argparse.ArgumentParser:
     train_util.add_sd_models_arguments(parser)
     train_util.add_dataset_arguments(parser, True, True, False)
     train_util.add_training_arguments(parser, True)
+    train_util.add_masked_loss_arguments(parser)
     train_util.add_optimizer_arguments(parser)
     config_util.add_config_arguments(parser)
     custom_train_functions.add_custom_train_arguments(parser, False)
diff --git a/train_textual_inversion_XTI.py b/train_textual_inversion_XTI.py
index 695fad2a8..72b79da46 100644
--- a/train_textual_inversion_XTI.py
+++ b/train_textual_inversion_XTI.py
@@ -9,6 +9,7 @@
 
 import torch
 from library.device_utils import init_ipex, clean_memory_on_device
+
 init_ipex()
 
 from accelerate.utils import set_seed
@@ -31,6 +32,7 @@
     apply_noise_offset,
     scale_v_prediction_loss_like_noise_prediction,
     apply_debiased_estimation,
+    apply_masked_loss,
 )
 import library.original_unet as original_unet
 from XTI_hijack import unet_forward_XTI, downblock_forward_XTI, upblock_forward_XTI
@@ -200,7 +202,7 @@ def train(args):
     logger.info(f"create embeddings for {args.num_vectors_per_token} tokens, for {args.token_string}")
 
     # データセットを準備する
-    blueprint_generator = BlueprintGenerator(ConfigSanitizer(True, True, False, False))
+    blueprint_generator = BlueprintGenerator(ConfigSanitizer(True, True, args.masked_loss, False))
     if args.dataset_config is not None:
         logger.info(f"Load dataset config from {args.dataset_config}")
         user_config = config_util.load_user_config(args.dataset_config)
@@ -471,6 +473,8 @@ def remove_model(old_ckpt_name):
                     target = noise
 
                 loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none")
+                if args.masked_loss:
+                    loss = apply_masked_loss(loss, batch)
                 loss = loss.mean([1, 2, 3])
 
                 loss_weights = batch["loss_weights"]  # 各sampleごとのweight
@@ -662,6 +666,7 @@ def setup_parser() -> argparse.ArgumentParser:
     train_util.add_sd_models_arguments(parser)
     train_util.add_dataset_arguments(parser, True, True, False)
     train_util.add_training_arguments(parser, True)
+    train_util.add_masked_loss_arguments(parser)
     train_util.add_optimizer_arguments(parser)
     config_util.add_config_arguments(parser)
     custom_train_functions.add_custom_train_arguments(parser, False)

From a7dff592d34a5dd9d306de822db70f0028676cab Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E9=9D=92=E9=BE=8D=E8=81=96=E8=80=85=40bdsqlsz?=
 <qinglongshengzhe@gmail.com>
Date: Mon, 18 Mar 2024 22:29:05 +0800
Subject: [PATCH 28/69] Update tag_images_by_wd14_tagger.py

add WDV3
---
 finetune/tag_images_by_wd14_tagger.py | 21 ++++++++++++---------
 1 file changed, 12 insertions(+), 9 deletions(-)

diff --git a/finetune/tag_images_by_wd14_tagger.py b/finetune/tag_images_by_wd14_tagger.py
index b56d921a3..e63ec3eb4 100644
--- a/finetune/tag_images_by_wd14_tagger.py
+++ b/finetune/tag_images_by_wd14_tagger.py
@@ -86,23 +86,26 @@ def main(args):
         logger.info(f"downloading wd14 tagger model from hf_hub. id: {args.repo_id}")
         files = FILES
         if args.onnx:
+            files = ["selected_tags.csv"]
             files += FILES_ONNX
+        else:
+            for file in SUB_DIR_FILES:
+                hf_hub_download(
+                    args.repo_id,
+                    file,
+                    subfolder=SUB_DIR,
+                    cache_dir=os.path.join(args.model_dir, SUB_DIR),
+                    force_download=True,
+                    force_filename=file,
+                )
         for file in files:
             hf_hub_download(args.repo_id, file, cache_dir=args.model_dir, force_download=True, force_filename=file)
-        for file in SUB_DIR_FILES:
-            hf_hub_download(
-                args.repo_id,
-                file,
-                subfolder=SUB_DIR,
-                cache_dir=os.path.join(args.model_dir, SUB_DIR),
-                force_download=True,
-                force_filename=file,
-            )
     else:
         logger.info("using existing wd14 tagger model")
 
     # 画像を読み込む
     if args.onnx:
+        import torch
         import onnx
         import onnxruntime as ort
 

From 5410a8c79b23c594bb340050b4a81e30d95cd7be Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E9=9D=92=E9=BE=8D=E8=81=96=E8=80=85=40bdsqlsz?=
 <qinglongshengzhe@gmail.com>
Date: Mon, 18 Mar 2024 22:31:00 +0800
Subject: [PATCH 29/69] Update requirements.txt

---
 requirements.txt | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/requirements.txt b/requirements.txt
index 279de350c..326b65b3e 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -22,9 +22,9 @@ huggingface-hub==0.20.1
 # for WD14 captioning (tensorflow)
 # tensorflow==2.10.1
 # for WD14 captioning (onnx)
-# onnx==1.14.1
-# onnxruntime-gpu==1.16.0
-# onnxruntime==1.16.0
+# onnx==1.15.1
+# onnxruntime-gpu==1.17.1
+# onnxruntime==1.17.1
 # this is for onnx: 
 # protobuf==3.20.3
 # open clip for SDXL

From a71c35ccd9c813821fcbd3f0e00d71fb5e6d91d6 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E9=9D=92=E9=BE=8D=E8=81=96=E8=80=85=40bdsqlsz?=
 <qinglongshengzhe@gmail.com>
Date: Mon, 18 Mar 2024 22:31:59 +0800
Subject: [PATCH 30/69] Update requirements.txt

---
 requirements.txt | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/requirements.txt b/requirements.txt
index 326b65b3e..6898eccf6 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -25,6 +25,9 @@ huggingface-hub==0.20.1
 # onnx==1.15.1
 # onnxruntime-gpu==1.17.1
 # onnxruntime==1.17.1
+# for cuda 12.1(default 11.8)
+# onnxruntime-gpu --extra-index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/onnxruntime-cuda-12/pypi/simple/
+
 # this is for onnx: 
 # protobuf==3.20.3
 # open clip for SDXL

From 6c51c971d135a346d2f9081760f138b1c6515e9b Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E9=9D=92=E9=BE=8D=E8=81=96=E8=80=85=40bdsqlsz?=
 <qinglongshengzhe@gmail.com>
Date: Wed, 20 Mar 2024 09:35:21 +0800
Subject: [PATCH 31/69] fix typo

---
 requirements.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/requirements.txt b/requirements.txt
index 6898eccf6..805f0501d 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -22,7 +22,7 @@ huggingface-hub==0.20.1
 # for WD14 captioning (tensorflow)
 # tensorflow==2.10.1
 # for WD14 captioning (onnx)
-# onnx==1.15.1
+# onnx==1.15.0
 # onnxruntime-gpu==1.17.1
 # onnxruntime==1.17.1
 # for cuda 12.1(default 11.8)

From 80dbbf5e4875f56ff1e0d8aacea4e73b96a14b63 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Wed, 20 Mar 2024 16:14:57 +0900
Subject: [PATCH 32/69] tagger now stores model under repo_id subdir

---
 README.md                             |  9 ++++-
 finetune/tag_images_by_wd14_tagger.py | 55 ++++++++++++++++++---------
 2 files changed, 45 insertions(+), 19 deletions(-)

diff --git a/README.md b/README.md
index f0cad611f..d03204037 100644
--- a/README.md
+++ b/README.md
@@ -260,7 +260,9 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
   - `keep_tokens_separator` is updated to be used twice in the caption. When you specify `keep_tokens_separator="|||"`, the part divided by the second `|||` is not shuffled or dropped and remains at the end.
   - The existing features `caption_prefix` and `caption_suffix` can be used together. `caption_prefix` and `caption_suffix` are processed first, and then `enable_wildcard`, `keep_tokens_separator`, shuffling and dropping, and `secondary_separator` are processed in order.
   - The examples are [shown below](#example-of-dataset-settings--データセット設定の記述例).
-
+- The support for v3 repositories is added to `tag_image_by_wd14_tagger.py` (`--onnx` option only). PR [#1192](https://github.com/kohya-ss/sd-scripts/pull/1192) Thanks to sdbds!
+  - Onnx may need to be updated. Onnx is not installed by default, so please install or update it with `pip install onnx==1.15.0 onnxruntime-gpu==1.17.1` etc. Please also check the comments in `requirements.txt`.
+- The model is now saved in the subdirectory as `--repo_id` in `tag_image_by_wd14_tagger.py` . This caches multiple repo_id models. Please delete unnecessary files under `--model_dir`.
 
 - Colab での動作時、ログ出力で停止してしまうようです。学習スクリプトに `--console_log_simple` オプションを指定し、rich のロギングを無効してお試しください。
 - `train_network.py` および `sdxl_train_network.py` で、学習したモデルのメタデータに一部のデータセット設定が記録されるよう修正しました（`caption_prefix`、`caption_suffix`、`keep_tokens_separator`、`secondary_separator`、`enable_wildcard`）。
@@ -269,6 +271,11 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
   - `enable_wildcard` を追加しました。`true` にするとワイルドカード記法 `{aaa|bbb|ccc}` が使えます。詳しくは記述例をご覧ください。
   - `keep_tokens_separator` をキャプション内に 2 つ使えるようにしました。たとえば `keep_tokens_separator="|||"` と指定したとき、`1girl, hatsune miku, vocaloid ||| stage, mic ||| best quality, rating: general` とキャプションを指定すると、二番目の `|||` で分割された部分はシャッフル、drop されず末尾に残ります。
   - 既存の機能 `caption_prefix` と `caption_suffix` とあわせて使えます。`caption_prefix` と `caption_suffix` は一番最初に処理され、その後、ワイルドカード、`keep_tokens_separator`、シャッフルおよび drop、`secondary_separator` の順に処理されます。
+- `tag_image_by_wd14_tagger.py` で v3 のリポジトリがサポートされました（`--onnx` 指定時のみ有効）。 PR [#1192](https://github.com/kohya-ss/sd-scripts/pull/1192) sdbds 氏に感謝します。
+  - Onnx のバージョンアップが必要になるかもしれません。デフォルトでは Onnx はインストールされていませんので、`pip install onnx==1.15.0 onnxruntime-gpu==1.17.1` 等でインストール、アップデートしてください。`requirements.txt` のコメントもあわせてご確認ください。
+- `tag_image_by_wd14_tagger.py` で、モデルを`--repo_id` のサブディレクトリに保存するようにしました。これにより複数のモデルファイルがキャッシュされます。`--model_dir` 直下の不要なファイルは削除願います。
+
+
 
 #### Example of dataset settings / データセット設定の記述例:
 
diff --git a/finetune/tag_images_by_wd14_tagger.py b/finetune/tag_images_by_wd14_tagger.py
index e63ec3eb4..401c6d1ec 100644
--- a/finetune/tag_images_by_wd14_tagger.py
+++ b/finetune/tag_images_by_wd14_tagger.py
@@ -12,8 +12,10 @@
 
 import library.train_util as train_util
 from library.utils import setup_logging
+
 setup_logging()
 import logging
+
 logger = logging.getLogger(__name__)
 
 # from wd14 tagger
@@ -79,10 +81,15 @@ def collate_fn_remove_corrupted(batch):
 
 
 def main(args):
+    # model location is model_dir + repo_id
+    # repo id may be like "user/repo" or "user/repo/branch", so we need to remove slash
+    model_location = os.path.join(args.model_dir, args.repo_id.replace("/", "_"))
+
     # hf_hub_downloadをそのまま使うとsymlink関係で問題があるらしいので、キャッシュディレクトリとforce_filenameを指定してなんとかする
     # depreacatedの警告が出るけどなくなったらその時
     # https://github.com/toriato/stable-diffusion-webui-wd14-tagger/issues/22
-    if not os.path.exists(args.model_dir) or args.force_download:
+    if not os.path.exists(model_location) or args.force_download:
+        os.makedirs(args.model_dir, exist_ok=True)
         logger.info(f"downloading wd14 tagger model from hf_hub. id: {args.repo_id}")
         files = FILES
         if args.onnx:
@@ -94,12 +101,12 @@ def main(args):
                     args.repo_id,
                     file,
                     subfolder=SUB_DIR,
-                    cache_dir=os.path.join(args.model_dir, SUB_DIR),
+                    cache_dir=os.path.join(model_location, SUB_DIR),
                     force_download=True,
                     force_filename=file,
                 )
         for file in files:
-            hf_hub_download(args.repo_id, file, cache_dir=args.model_dir, force_download=True, force_filename=file)
+            hf_hub_download(args.repo_id, file, cache_dir=model_location, force_download=True, force_filename=file)
     else:
         logger.info("using existing wd14 tagger model")
 
@@ -109,7 +116,7 @@ def main(args):
         import onnx
         import onnxruntime as ort
 
-        onnx_path = f"{args.model_dir}/model.onnx"
+        onnx_path = f"{model_location}/model.onnx"
         logger.info("Running wd14 tagger with onnx")
         logger.info(f"loading onnx model: {onnx_path}")
 
@@ -126,7 +133,7 @@ def main(args):
         except:
             batch_size = model.graph.input[0].type.tensor_type.shape.dim[0].dim_param
 
-        if args.batch_size != batch_size and type(batch_size) != str:
+        if args.batch_size != batch_size and type(batch_size) != str and batch_size > 0:
             # some rebatch model may use 'N' as dynamic axes
             logger.warning(
                 f"Batch size {args.batch_size} doesn't match onnx model batch size {batch_size}, use model batch size {batch_size}"
@@ -137,19 +144,19 @@ def main(args):
 
         ort_sess = ort.InferenceSession(
             onnx_path,
-            providers=["CUDAExecutionProvider"]
-            if "CUDAExecutionProvider" in ort.get_available_providers()
-            else ["CPUExecutionProvider"],
+            providers=(
+                ["CUDAExecutionProvider"] if "CUDAExecutionProvider" in ort.get_available_providers() else ["CPUExecutionProvider"]
+            ),
         )
     else:
         from tensorflow.keras.models import load_model
 
-        model = load_model(f"{args.model_dir}")
+        model = load_model(f"{model_location}")
 
     # label_names = pd.read_csv("2022_0000_0899_6549/selected_tags.csv")
     # 依存ライブラリを増やしたくないので自力で読むよ
 
-    with open(os.path.join(args.model_dir, CSV_FILE), "r", encoding="utf-8") as f:
+    with open(os.path.join(model_location, CSV_FILE), "r", encoding="utf-8") as f:
         reader = csv.reader(f)
         l = [row for row in reader]
         header = l[0]  # tag_id,name,category,count
@@ -175,8 +182,8 @@ def run_batch(path_imgs):
         imgs = np.array([im for _, im in path_imgs])
 
         if args.onnx:
-            if len(imgs) < args.batch_size:
-                imgs = np.concatenate([imgs, np.zeros((args.batch_size - len(imgs), IMAGE_SIZE, IMAGE_SIZE, 3))], axis=0)
+            # if len(imgs) < args.batch_size:
+            #     imgs = np.concatenate([imgs, np.zeros((args.batch_size - len(imgs), IMAGE_SIZE, IMAGE_SIZE, 3))], axis=0)
             probs = ort_sess.run(None, {input_name: imgs})[0]  # onnx output numpy
             probs = probs[: len(path_imgs)]
         else:
@@ -317,7 +324,9 @@ def setup_parser() -> argparse.ArgumentParser:
         help="directory to store wd14 tagger model / wd14 taggerのモデルを格納するディレクトリ",
     )
     parser.add_argument(
-        "--force_download", action="store_true", help="force downloading wd14 tagger models / wd14 taggerのモデルを再ダウンロードします"
+        "--force_download",
+        action="store_true",
+        help="force downloading wd14 tagger models / wd14 taggerのモデルを再ダウンロードします",
     )
     parser.add_argument("--batch_size", type=int, default=1, help="batch size in inference / 推論時のバッチサイズ")
     parser.add_argument(
@@ -332,8 +341,12 @@ def setup_parser() -> argparse.ArgumentParser:
         default=None,
         help="extension of caption file (for backward compatibility) / 出力されるキャプションファイルの拡張子（スペルミスしていたのを残してあります）",
     )
-    parser.add_argument("--caption_extension", type=str, default=".txt", help="extension of caption file / 出力されるキャプションファイルの拡張子")
-    parser.add_argument("--thresh", type=float, default=0.35, help="threshold of confidence to add a tag / タグを追加するか判定する閾値")
+    parser.add_argument(
+        "--caption_extension", type=str, default=".txt", help="extension of caption file / 出力されるキャプションファイルの拡張子"
+    )
+    parser.add_argument(
+        "--thresh", type=float, default=0.35, help="threshold of confidence to add a tag / タグを追加するか判定する閾値"
+    )
     parser.add_argument(
         "--general_threshold",
         type=float,
@@ -346,7 +359,9 @@ def setup_parser() -> argparse.ArgumentParser:
         default=None,
         help="threshold of confidence to add a tag for character category, same as --thres if omitted / characterカテゴリのタグを追加するための確信度の閾値、省略時は --thresh と同じ",
     )
-    parser.add_argument("--recursive", action="store_true", help="search for images in subfolders recursively / サブフォルダを再帰的に検索する")
+    parser.add_argument(
+        "--recursive", action="store_true", help="search for images in subfolders recursively / サブフォルダを再帰的に検索する"
+    )
     parser.add_argument(
         "--remove_underscore",
         action="store_true",
@@ -359,9 +374,13 @@ def setup_parser() -> argparse.ArgumentParser:
         default="",
         help="comma-separated list of undesired tags to remove from the output / 出力から除外したいタグのカンマ区切りのリスト",
     )
-    parser.add_argument("--frequency_tags", action="store_true", help="Show frequency of tags for images / 画像ごとのタグの出現頻度を表示する")
+    parser.add_argument(
+        "--frequency_tags", action="store_true", help="Show frequency of tags for images / 画像ごとのタグの出現頻度を表示する"
+    )
     parser.add_argument("--onnx", action="store_true", help="use onnx model for inference / onnxモデルを推論に使用する")
-    parser.add_argument("--append_tags", action="store_true", help="Append captions instead of overwriting / 上書きではなくキャプションを追記する")
+    parser.add_argument(
+        "--append_tags", action="store_true", help="Append captions instead of overwriting / 上書きではなくキャプションを追記する"
+    )
     parser.add_argument(
         "--caption_separator",
         type=str,

From 46331a9e8ef695ea0b5a19686202d011109a56b6 Mon Sep 17 00:00:00 2001
From: Victor Espinoza-Guerra <thesilvervelvets@gmail.com>
Date: Wed, 20 Mar 2024 00:31:01 -0700
Subject: [PATCH 33/69] English Translation of config_README-ja.md (#1175)

* Add files via upload

Creating template to work on.

* Update config_README-en.md

Total Conversion from Japanese to English.

* Update config_README-en.md

* Update config_README-en.md

* Update config_README-en.md
---
 docs/config_README-en.md | 279 +++++++++++++++++++++++++++++++++++++++
 1 file changed, 279 insertions(+)
 create mode 100644 docs/config_README-en.md

diff --git a/docs/config_README-en.md b/docs/config_README-en.md
new file mode 100644
index 000000000..a0727934d
--- /dev/null
+++ b/docs/config_README-en.md
@@ -0,0 +1,279 @@
+Original Source by kohya-ss
+
+A.I Translation by Model: NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO, editing by Darkstorm2150
+
+# Config Readme
+
+This README is about the configuration files that can be passed with the `--dataset_config` option.
+
+## Overview
+
+By passing a configuration file, users can make detailed settings.
+
+* Multiple datasets can be configured
+   * For example, by setting `resolution` for each dataset, they can be mixed and trained.
+   * In training methods that support both the DreamBooth approach and the fine-tuning approach, datasets of the DreamBooth method and the fine-tuning method can be mixed.
+* Settings can be changed for each subset
+   * A subset is a partition of the dataset by image directory or metadata. Several subsets make up a dataset.
+   * Options such as `keep_tokens` and `flip_aug` can be set for each subset. On the other hand, options such as `resolution` and `batch_size` can be set for each dataset, and their values are common among subsets belonging to the same dataset. More details will be provided later.
+
+The configuration file format can be JSON or TOML. Considering the ease of writing, it is recommended to use [TOML](https://toml.io/ja/v1.0.0-rc.2). The following explanation assumes the use of TOML.
+
+
+Here is an example of a configuration file written in TOML.
+
+```toml
+[general]
+shuffle_caption = true
+caption_extension = '.txt'
+keep_tokens = 1
+
+# This is a DreamBooth-style dataset
+[[datasets]]
+resolution = 512
+batch_size = 4
+keep_tokens = 2
+
+  [[datasets.subsets]]
+  image_dir = 'C:\hoge'
+  class_tokens = 'hoge girl'
+  # This subset uses keep_tokens = 2 (the value of the parent datasets)
+
+  [[datasets.subsets]]
+  image_dir = 'C:\fuga'
+  class_tokens = 'fuga boy'
+  keep_tokens = 3
+
+  [[datasets.subsets]]
+  is_reg = true
+  image_dir = 'C:\reg'
+  class_tokens = 'human'
+  keep_tokens = 1
+
+# This is a fine-tuning dataset
+[[datasets]]
+resolution = [768, 768]
+batch_size = 2
+
+  [[datasets.subsets]]
+  image_dir = 'C:\piyo'
+  metadata_file = 'C:\piyo\piyo_md.json'
+  # This subset uses keep_tokens = 1 (the value of [general])
+```
+
+In this example, three directories are trained as a DreamBooth-style dataset at 512x512 (batch size 4), and one directory is trained as a fine-tuning dataset at 768x768 (batch size 2).
+
+## Settings for datasets and subsets
+
+Settings for datasets and subsets are divided into several registration locations.
+
+* `[general]`
+    * This is where options that apply to all datasets or all subsets are specified.
+    * If there are options with the same name in the dataset-specific or subset-specific settings, the dataset-specific or subset-specific settings take precedence.
+* `[[datasets]]`
+    * `datasets` is where settings for datasets are registered. This is where options that apply individually to each dataset are specified.
+	* If there are subset-specific settings, the subset-specific settings take precedence.
+* `[[datasets.subsets]]`
+    * `datasets.subsets` is where settings for subsets are registered. This is where options that apply individually to each subset are specified.
+
+Here is an image showing the correspondence between image directories and registration locations in the previous example.
+
+```
+C:\
+├─ hoge  ->  [[datasets.subsets]] No.1  ┐                        ┐
+├─ fuga  ->  [[datasets.subsets]] No.2  |->  [[datasets]] No.1   |->  [general]
+├─ reg   ->  [[datasets.subsets]] No.3  ┘                        |
+└─ piyo  ->  [[datasets.subsets]] No.4  -->  [[datasets]] No.2   ┘
+```
+
+The image directory corresponds to each `[[datasets.subsets]]`. Then, multiple `[[datasets.subsets]]` are combined to form one `[[datasets]]`. All `[[datasets]]` and `[[datasets.subsets]]` belong to `[general]`.
+
+The available options for each registration location may differ, but if the same option is specified, the value in the lower registration location will take precedence. You can check how the `keep_tokens` option is handled in the previous example for better understanding.
+
+Additionally, the available options may vary depending on the method that the learning approach supports.
+
+* Options specific to the DreamBooth method
+* Options specific to the fine-tuning method
+* Options available when using the caption dropout technique
+
+When using both the DreamBooth method and the fine-tuning method, they can be used together with a learning approach that supports both.
+When using them together, a point to note is that the method is determined based on the dataset, so it is not possible to mix DreamBooth method subsets and fine-tuning method subsets within the same dataset.
+In other words, if you want to use both methods together, you need to set up subsets of different methods belonging to different datasets.
+
+In terms of program behavior, if the `metadata_file` option exists, it is determined to be a subset of fine-tuning. Therefore, for subsets belonging to the same dataset, as long as they are either "all have the `metadata_file` option" or "all have no `metadata_file` option," there is no problem.
+
+Below, the available options will be explained. For options with the same name as the command-line argument, the explanation will be omitted in principle. Please refer to other READMEs.
+
+### Common options for all learning methods
+
+These are options that can be specified regardless of the learning method.
+
+#### Data set specific options
+
+These are options related to the configuration of the data set. They cannot be described in `datasets.subsets`.
+
+
+| Option Name | Example Setting | `[general]` | `[[datasets]]` |
+| ---- | ---- | ---- | ---- |
+| `batch_size` | `1` | o | o |
+| `bucket_no_upscale` | `true` | o | o |
+| `bucket_reso_steps` | `64` | o | o |
+| `enable_bucket` | `true` | o | o |
+| `max_bucket_reso` | `1024` | o | o |
+| `min_bucket_reso` | `128` | o | o |
+| `resolution` | `256`, `[512, 512]` | o | o |
+
+* `batch_size`
+    * This corresponds to the command-line argument `--train_batch_size`.
+
+These settings are fixed per dataset. That means that subsets belonging to the same dataset will share these settings. For example, if you want to prepare datasets with different resolutions, you can define them as separate datasets as shown in the example above, and set different resolutions for each.
+
+#### Options for Subsets
+
+These options are related to subset configuration.
+
+| Option Name | Example | `[general]` | `[[datasets]]` | `[[dataset.subsets]]` |
+| ---- | ---- | ---- | ---- | ---- |
+| `color_aug` | `false` | o | o | o |
+| `face_crop_aug_range` | `[1.0, 3.0]` | o | o | o |
+| `flip_aug` | `true` | o | o | o |
+| `keep_tokens` | `2` | o | o | o |
+| `num_repeats` | `10` | o | o | o |
+| `random_crop` | `false` | o | o | o |
+| `shuffle_caption` | `true` | o | o | o |
+| `caption_prefix` | `"masterpiece, best quality, "` | o | o | o |
+| `caption_suffix` | `", from side"` | o | o | o |
+
+* `num_repeats`
+    * Specifies the number of repeats for images in a subset. This is equivalent to `--dataset_repeats` in fine-tuning but can be specified for any training method.
+* `caption_prefix`, `caption_suffix`
+    * Specifies the prefix and suffix strings to be appended to the captions. Shuffling is performed with these strings included. Be cautious when using `keep_tokens`.
+
+### DreamBooth-specific options
+
+DreamBooth-specific options only exist as subsets-specific options.
+
+#### Subset-specific options
+
+Options related to the configuration of DreamBooth subsets.
+
+| Option Name | Example Setting | `[general]` | `[[datasets]]` | `[[dataset.subsets]]` |
+| ---- | ---- | ---- | ---- | ---- |
+| `image_dir` | `'C:\hoge'` | - | - | o (required) |
+| `caption_extension` | `".txt"` | o | o | o |
+| `class_tokens` | `"sks girl"` | - | - | o |
+| `is_reg` | `false` | - | - | o |
+
+Firstly, note that for `image_dir`, the path to the image files must be specified as being directly in the directory. Unlike the previous DreamBooth method, where images had to be placed in subdirectories, this is not compatible with that specification. Also, even if you name the folder something like "5_cat", the number of repeats of the image and the class name will not be reflected. If you want to set these individually, you will need to explicitly specify them using `num_repeats` and `class_tokens`.
+
+* `image_dir`
+    * Specifies the path to the image directory. This is a required option.
+    * Images must be placed directly under the directory.
+* `class_tokens`
+    * Sets the class tokens.
+    * Only used during training when a corresponding caption file does not exist. The determination of whether or not to use it is made on a per-image basis. If `class_tokens` is not specified and a caption file is not found, an error will occur.
+* `is_reg`
+    * Specifies whether the subset images are for normalization. If not specified, it is set to `false`, meaning that the images are not for normalization.
+
+### Fine-tuning method specific options
+
+The options for the fine-tuning method only exist for subset-specific options.
+
+#### Subset-specific options
+
+These options are related to the configuration of the fine-tuning method's subsets.
+
+| Option name | Example setting | `[general]` | `[[datasets]]` | `[[dataset.subsets]]` |
+| ---- | ---- | ---- | ---- | ---- |
+| `image_dir` | `'C:\hoge'` | - | - | o |
+| `metadata_file` | `'C:\piyo\piyo_md.json'` | - | - | o (required) |
+
+* `image_dir`
+    * Specify the path to the image directory. Unlike the DreamBooth method, specifying it is not mandatory, but it is recommended to do so.
+        * The case where it is not necessary to specify is when the `--full_path` is added to the command line when generating the metadata file.
+    * The images must be placed directly under the directory.
+* `metadata_file`
+    * Specify the path to the metadata file used for the subset. This is a required option.
+        * It is equivalent to the command-line argument `--in_json`.
+    * Due to the specification that a metadata file must be specified for each subset, it is recommended to avoid creating a metadata file with images from different directories as a single metadata file. It is strongly recommended to prepare a separate metadata file for each image directory and register them as separate subsets.
+
+### Options available when caption dropout method can be used
+
+The options available when the caption dropout method can be used exist only for subsets. Regardless of whether it's the DreamBooth method or fine-tuning method, if it supports caption dropout, it can be specified.
+
+#### Subset-specific options
+
+Options related to the setting of subsets that caption dropout can be used for.
+
+| Option Name | `[general]` | `[[datasets]]` | `[[dataset.subsets]]` |
+| ---- | ---- | ---- | ---- |
+| `caption_dropout_every_n_epochs` | o | o | o |
+| `caption_dropout_rate` | o | o | o |
+| `caption_tag_dropout_rate` | o | o | o |
+
+## Behavior when there are duplicate subsets
+
+In the case of the DreamBooth dataset, if there are multiple `image_dir` directories with the same content, they are considered to be duplicate subsets. For the fine-tuning dataset, if there are multiple `metadata_file` files with the same content, they are considered to be duplicate subsets. If duplicate subsets exist in the dataset, subsequent subsets will be ignored.
+
+However, if they belong to different datasets, they are not considered duplicates. For example, if you have subsets with the same `image_dir` in different datasets, they will not be considered duplicates. This is useful when you want to train with the same image but with different resolutions.
+
+```toml
+# If data sets exist separately, they are not considered duplicates and are both used for training.
+
+[[datasets]]
+resolution = 512
+
+  [[datasets.subsets]]
+  image_dir = 'C:\hoge'
+
+[[datasets]]
+resolution = 768
+
+  [[datasets.subsets]]
+  image_dir = 'C:\hoge'
+```
+
+## Command Line Argument and Configuration File
+
+There are options in the configuration file that have overlapping roles with command line argument options.
+
+The following command line argument options are ignored if a configuration file is passed:
+
+* `--train_data_dir`
+* `--reg_data_dir`
+* `--in_json`
+
+The following command line argument options are given priority over the configuration file options if both are specified simultaneously. In most cases, they have the same names as the corresponding options in the configuration file.
+
+| Command Line Argument Option   | Prioritized Configuration File Option |
+| ------------------------------- | ------------------------------------- |
+| `--bucket_no_upscale`           |                                       |
+| `--bucket_reso_steps`           |                                       |
+| `--caption_dropout_every_n_epochs` |                                       |
+| `--caption_dropout_rate`        |                                       |
+| `--caption_extension`           |                                       |
+| `--caption_tag_dropout_rate`    |                                       |
+| `--color_aug`                   |                                       |
+| `--dataset_repeats`             | `num_repeats`                          |
+| `--enable_bucket`               |                                       |
+| `--face_crop_aug_range`         |                                       |
+| `--flip_aug`                    |                                       |
+| `--keep_tokens`                 |                                       |
+| `--min_bucket_reso`              |                                       |
+| `--random_crop`                 |                                       |
+| `--resolution`                  |                                       |
+| `--shuffle_caption`             |                                       |
+| `--train_batch_size`            | `batch_size`                           |
+
+## Error Guide
+
+Currently, we are using an external library to check if the configuration file is written correctly, but the development has not been completed, and there is a problem that the error message is not clear. In the future, we plan to improve this problem.
+
+As a temporary measure, we will list common errors and their solutions. If you encounter an error even though it should be correct or if the error content is not understandable, please contact us as it may be a bug.
+
+* `voluptuous.error.MultipleInvalid: required key not provided @ ...`: This error occurs when a required option is not provided. It is highly likely that you forgot to specify the option or misspelled the option name.
+  * The error location is indicated by `...` in the error message. For example, if you encounter an error like `voluptuous.error.MultipleInvalid: required key not provided @ data['datasets'][0]['subsets'][0]['image_dir']`, it means that the `image_dir` option does not exist in the 0th `subsets` of the 0th `datasets` setting.
+* `voluptuous.error.MultipleInvalid: expected int for dictionary value @ ...`: This error occurs when the specified value format is incorrect. It is highly likely that the value format is incorrect. The `int` part changes depending on the target option. The example configurations in this README may be helpful.
+* `voluptuous.error.MultipleInvalid: extra keys not allowed @ ...`: This error occurs when there is an option name that is not supported. It is highly likely that you misspelled the option name or mistakenly included it.
+
+

From 5f6196e4c71763250da316cc0f4ce15db1696017 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Wed, 20 Mar 2024 16:35:23 +0900
Subject: [PATCH 34/69] update readme

---
 README.md | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/README.md b/README.md
index d03204037..f8601c1b2 100644
--- a/README.md
+++ b/README.md
@@ -33,6 +33,7 @@ Most of the documents are written in Japanese.
 * [Training guide - common](./docs/train_README-ja.md) : data preparation, options etc... 
   * [Chinese version](./docs/train_README-zh.md)
 * [Dataset config](./docs/config_README-ja.md) 
+  * [English version](./docs/config_README-en.md)
 * [DreamBooth training guide](./docs/train_db_README-ja.md)
 * [Step by Step fine-tuning guide](./docs/fine_tune_README_ja.md):
 * [training LoRA](./docs/train_network_README-ja.md)
@@ -263,6 +264,8 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
 - The support for v3 repositories is added to `tag_image_by_wd14_tagger.py` (`--onnx` option only). PR [#1192](https://github.com/kohya-ss/sd-scripts/pull/1192) Thanks to sdbds!
   - Onnx may need to be updated. Onnx is not installed by default, so please install or update it with `pip install onnx==1.15.0 onnxruntime-gpu==1.17.1` etc. Please also check the comments in `requirements.txt`.
 - The model is now saved in the subdirectory as `--repo_id` in `tag_image_by_wd14_tagger.py` . This caches multiple repo_id models. Please delete unnecessary files under `--model_dir`.
+- The options `--noise_offset_random_strength` and `--ip_noise_gamma_random_strength` are added to each training script. These options can be used to vary the noise offset and ip noise gamma in the range of 0 to the specified value. PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) Thanks to KohakuBlueleaf!
+- The [English version of the dataset settings documentation](./docs/config_README-en.md) is added. PR [#1175](https://github.com/kohya-ss/sd-scripts/pull/1175) Thanks to darkstorm2150!
 
 - Colab での動作時、ログ出力で停止してしまうようです。学習スクリプトに `--console_log_simple` オプションを指定し、rich のロギングを無効してお試しください。
 - `train_network.py` および `sdxl_train_network.py` で、学習したモデルのメタデータに一部のデータセット設定が記録されるよう修正しました（`caption_prefix`、`caption_suffix`、`keep_tokens_separator`、`secondary_separator`、`enable_wildcard`）。
@@ -274,7 +277,8 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
 - `tag_image_by_wd14_tagger.py` で v3 のリポジトリがサポートされました（`--onnx` 指定時のみ有効）。 PR [#1192](https://github.com/kohya-ss/sd-scripts/pull/1192) sdbds 氏に感謝します。
   - Onnx のバージョンアップが必要になるかもしれません。デフォルトでは Onnx はインストールされていませんので、`pip install onnx==1.15.0 onnxruntime-gpu==1.17.1` 等でインストール、アップデートしてください。`requirements.txt` のコメントもあわせてご確認ください。
 - `tag_image_by_wd14_tagger.py` で、モデルを`--repo_id` のサブディレクトリに保存するようにしました。これにより複数のモデルファイルがキャッシュされます。`--model_dir` 直下の不要なファイルは削除願います。
-
+- 各学習スクリプトに、noise offset、ip noise gammaを、それぞれ 0~指定した値の範囲で変動させるオプション `--noise_offset_random_strength` および `--ip_noise_gamma_random_strength` が追加されました。 PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) KohakuBlueleaf 氏に感謝します。
+- データセット設定の[英語版ドキュメント](./docs/config_README-en.md) が追加されました。PR [#1175](https://github.com/kohya-ss/sd-scripts/pull/1175) darkstorm2150 氏に感謝します。
 
 
 #### Example of dataset settings / データセット設定の記述例:

From 3b0db0f17f46148abe345c5cdce76ff707bdccd3 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Wed, 20 Mar 2024 17:45:35 +0900
Subject: [PATCH 35/69] update readme

---
 README.md | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/README.md b/README.md
index f8601c1b2..81c176a79 100644
--- a/README.md
+++ b/README.md
@@ -266,6 +266,8 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
 - The model is now saved in the subdirectory as `--repo_id` in `tag_image_by_wd14_tagger.py` . This caches multiple repo_id models. Please delete unnecessary files under `--model_dir`.
 - The options `--noise_offset_random_strength` and `--ip_noise_gamma_random_strength` are added to each training script. These options can be used to vary the noise offset and ip noise gamma in the range of 0 to the specified value. PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) Thanks to KohakuBlueleaf!
 - The [English version of the dataset settings documentation](./docs/config_README-en.md) is added. PR [#1175](https://github.com/kohya-ss/sd-scripts/pull/1175) Thanks to darkstorm2150!
+- The `.toml` file for the dataset config is now read in UTF-8 encoding. PR [#1167](https://github.com/kohya-ss/sd-scripts/pull/1167) Thanks to Horizon1704!
+
 
 - Colab での動作時、ログ出力で停止してしまうようです。学習スクリプトに `--console_log_simple` オプションを指定し、rich のロギングを無効してお試しください。
 - `train_network.py` および `sdxl_train_network.py` で、学習したモデルのメタデータに一部のデータセット設定が記録されるよう修正しました（`caption_prefix`、`caption_suffix`、`keep_tokens_separator`、`secondary_separator`、`enable_wildcard`）。
@@ -279,7 +281,7 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
 - `tag_image_by_wd14_tagger.py` で、モデルを`--repo_id` のサブディレクトリに保存するようにしました。これにより複数のモデルファイルがキャッシュされます。`--model_dir` 直下の不要なファイルは削除願います。
 - 各学習スクリプトに、noise offset、ip noise gammaを、それぞれ 0~指定した値の範囲で変動させるオプション `--noise_offset_random_strength` および `--ip_noise_gamma_random_strength` が追加されました。 PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) KohakuBlueleaf 氏に感謝します。
 - データセット設定の[英語版ドキュメント](./docs/config_README-en.md) が追加されました。PR [#1175](https://github.com/kohya-ss/sd-scripts/pull/1175) darkstorm2150 氏に感謝します。
-
+- データセット設定の `.toml` ファイルが UTF-8 encoding で読み込まれるようになりました。PR [#1167](https://github.com/kohya-ss/sd-scripts/pull/1167) Horizon1704 氏に感謝します。
 
 #### Example of dataset settings / データセット設定の記述例:
 

From 855add067b06464eaa47ed55840da0f17d675762 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Wed, 20 Mar 2024 18:14:05 +0900
Subject: [PATCH 36/69] update option help and readme

---
 README.md             | 7 +++++--
 library/train_util.py | 8 ++++----
 2 files changed, 9 insertions(+), 6 deletions(-)

diff --git a/README.md b/README.md
index 81c176a79..804bad84a 100644
--- a/README.md
+++ b/README.md
@@ -253,6 +253,7 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
 ### Working in progress
 
 - Colab seems to stop with log output. Try specifying `--console_log_simple` option in the training script to disable rich logging.
+- The `.toml` file for the dataset config is now read in UTF-8 encoding. PR [#1167](https://github.com/kohya-ss/sd-scripts/pull/1167) Thanks to Horizon1704!
 - `train_network.py` and `sdxl_train_network.py` are modified to record some dataset settings in the metadata of the trained model (`caption_prefix`, `caption_suffix`, `keep_tokens_separator`, `secondary_separator`, `enable_wildcard`).
 - Some features are added to the dataset subset settings.
   - `secondary_separator` is added to specify the tag separator that is not the target of shuffling or dropping. 
@@ -266,10 +267,11 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
 - The model is now saved in the subdirectory as `--repo_id` in `tag_image_by_wd14_tagger.py` . This caches multiple repo_id models. Please delete unnecessary files under `--model_dir`.
 - The options `--noise_offset_random_strength` and `--ip_noise_gamma_random_strength` are added to each training script. These options can be used to vary the noise offset and ip noise gamma in the range of 0 to the specified value. PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) Thanks to KohakuBlueleaf!
 - The [English version of the dataset settings documentation](./docs/config_README-en.md) is added. PR [#1175](https://github.com/kohya-ss/sd-scripts/pull/1175) Thanks to darkstorm2150!
-- The `.toml` file for the dataset config is now read in UTF-8 encoding. PR [#1167](https://github.com/kohya-ss/sd-scripts/pull/1167) Thanks to Horizon1704!
+- The options `--save_state_on_train_end` are added to each training script. PR [#1168](https://github.com/kohya-ss/sd-scripts/pull/1168) Thanks to gesen2egee!
 
 
 - Colab での動作時、ログ出力で停止してしまうようです。学習スクリプトに `--console_log_simple` オプションを指定し、rich のロギングを無効してお試しください。
+- データセット設定の `.toml` ファイルが UTF-8 encoding で読み込まれるようになりました。PR [#1167](https://github.com/kohya-ss/sd-scripts/pull/1167) Horizon1704 氏に感謝します。
 - `train_network.py` および `sdxl_train_network.py` で、学習したモデルのメタデータに一部のデータセット設定が記録されるよう修正しました（`caption_prefix`、`caption_suffix`、`keep_tokens_separator`、`secondary_separator`、`enable_wildcard`）。
 - データセットのサブセット設定にいくつかの機能を追加しました。
   - シャッフルの対象とならないタグ分割識別子の指定 `secondary_separator` を追加しました。`secondary_separator=";;;"` のように指定します。`secondary_separator` で区切ることで、その部分はシャッフル、drop 時にまとめて扱われます。詳しくは記述例をご覧ください。
@@ -281,7 +283,8 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
 - `tag_image_by_wd14_tagger.py` で、モデルを`--repo_id` のサブディレクトリに保存するようにしました。これにより複数のモデルファイルがキャッシュされます。`--model_dir` 直下の不要なファイルは削除願います。
 - 各学習スクリプトに、noise offset、ip noise gammaを、それぞれ 0~指定した値の範囲で変動させるオプション `--noise_offset_random_strength` および `--ip_noise_gamma_random_strength` が追加されました。 PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) KohakuBlueleaf 氏に感謝します。
 - データセット設定の[英語版ドキュメント](./docs/config_README-en.md) が追加されました。PR [#1175](https://github.com/kohya-ss/sd-scripts/pull/1175) darkstorm2150 氏に感謝します。
-- データセット設定の `.toml` ファイルが UTF-8 encoding で読み込まれるようになりました。PR [#1167](https://github.com/kohya-ss/sd-scripts/pull/1167) Horizon1704 氏に感謝します。
+- 各学習スクリプトに、学習終了時に state を保存する `--save_state_on_train_end` オプションが追加されました。 PR [#1168](https://github.com/kohya-ss/sd-scripts/pull/1168) gesen2egee 氏に感謝します。
+
 
 #### Example of dataset settings / データセット設定の記述例:
 
diff --git a/library/train_util.py b/library/train_util.py
index 23961505f..a13985ee2 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -2936,13 +2936,13 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
     parser.add_argument(
         "--save_state",
         action="store_true",
-        help="save training state additionally (including optimizer states etc.) / optimizerなど学習状態も含めたstateを追加で保存する",
+        help="save training state additionally (including optimizer states etc.) when saving model / optimizerなど学習状態も含めたstateをモデル保存時に追加で保存する",
     )
     parser.add_argument(
         "--save_state_on_train_end",
         action="store_true",
-        help="save training state additionally (including optimizer states etc.) on train end / optimizerなど学習状態も含めたstateを追加で保存する",
-    )   
+        help="save training state (including optimizer states etc.) on train end / optimizerなど学習状態も含めたstateを学習完了時に保存する",
+    )
     parser.add_argument("--resume", type=str, default=None, help="saved state to resume training / 学習再開するモデルのstate")
 
     parser.add_argument("--train_batch_size", type=int, default=1, help="batch size for training / 学習時のバッチサイズ")
@@ -3550,7 +3550,7 @@ def read_config_from_file(args: argparse.Namespace, parser: argparse.ArgumentPar
         exit(1)
 
     logger.info(f"Loading settings from {config_path}...")
-    with open(config_path, "r", encoding='utf-8') as f:
+    with open(config_path, "r", encoding="utf-8") as f:
         config_dict = toml.load(f)
 
     # combine all sections into one

From d9456020d7547743c809a7c93f9a487276a66c74 Mon Sep 17 00:00:00 2001
From: BootsofLagrangian <hard2251@yonsei.ac.kr>
Date: Wed, 20 Mar 2024 20:52:59 +0900
Subject: [PATCH 37/69] Fix most of ZeRO stage uses optimizer partitioning  -
 we have to prepare optimizer and ds_model at the same time.  -
 pull/1139#issuecomment-1986790007

Signed-off-by: BootsofLagrangian <hard2251@yonsei.ac.kr>
---
 sdxl_train.py | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)

diff --git a/sdxl_train.py b/sdxl_train.py
index 613fe30b3..2cb80b6b3 100644
--- a/sdxl_train.py
+++ b/sdxl_train.py
@@ -412,7 +412,10 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
             text_encoder1=text_encoder1 if train_text_encoder1 else None,
             text_encoder2=text_encoder2 if train_text_encoder2 else None,
         )
-        ds_model = accelerator.prepare(ds_model)
+        # most of ZeRO stage uses optimizer partitioning, so we have to prepare optimizer and ds_model at the same time. # pull/1139#issuecomment-1986790007
+        ds_model, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+            ds_model, optimizer, train_dataloader, lr_scheduler
+        )
         training_models = [ds_model]
 
     else:
@@ -423,8 +426,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
             text_encoder1 = accelerator.prepare(text_encoder1)
         if train_text_encoder2:
             text_encoder2 = accelerator.prepare(text_encoder2)
-
-    optimizer, train_dataloader, lr_scheduler = accelerator.prepare(optimizer, train_dataloader, lr_scheduler)
+        optimizer, train_dataloader, lr_scheduler = accelerator.prepare(optimizer, train_dataloader, lr_scheduler)
 
     # TextEncoderの出力をキャッシュするときにはCPUへ移動する
     if args.cache_text_encoder_outputs:

From d17c0f508416d734360393804732bfa420fe1c27 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Thu, 21 Mar 2024 08:31:29 +0900
Subject: [PATCH 38/69] update dataset config doc

---
 README.md                | 88 +---------------------------------------
 docs/config_README-en.md | 73 +++++++++++++++++++++++++++++++++
 docs/config_README-ja.md | 75 +++++++++++++++++++++++++++++++++-
 3 files changed, 148 insertions(+), 88 deletions(-)

diff --git a/README.md b/README.md
index 804bad84a..dae311325 100644
--- a/README.md
+++ b/README.md
@@ -261,7 +261,7 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
   - `enable_wildcard` is added. When set to `true`, the wildcard notation `{aaa|bbb|ccc}` can be used. See the example below.
   - `keep_tokens_separator` is updated to be used twice in the caption. When you specify `keep_tokens_separator="|||"`, the part divided by the second `|||` is not shuffled or dropped and remains at the end.
   - The existing features `caption_prefix` and `caption_suffix` can be used together. `caption_prefix` and `caption_suffix` are processed first, and then `enable_wildcard`, `keep_tokens_separator`, shuffling and dropping, and `secondary_separator` are processed in order.
-  - The examples are [shown below](#example-of-dataset-settings--データセット設定の記述例).
+  - See [Dataset config](./docs/config_README-en.md) for details.
 - The support for v3 repositories is added to `tag_image_by_wd14_tagger.py` (`--onnx` option only). PR [#1192](https://github.com/kohya-ss/sd-scripts/pull/1192) Thanks to sdbds!
   - Onnx may need to be updated. Onnx is not installed by default, so please install or update it with `pip install onnx==1.15.0 onnxruntime-gpu==1.17.1` etc. Please also check the comments in `requirements.txt`.
 - The model is now saved in the subdirectory as `--repo_id` in `tag_image_by_wd14_tagger.py` . This caches multiple repo_id models. Please delete unnecessary files under `--model_dir`.
@@ -278,6 +278,7 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
   - `enable_wildcard` を追加しました。`true` にするとワイルドカード記法 `{aaa|bbb|ccc}` が使えます。詳しくは記述例をご覧ください。
   - `keep_tokens_separator` をキャプション内に 2 つ使えるようにしました。たとえば `keep_tokens_separator="|||"` と指定したとき、`1girl, hatsune miku, vocaloid ||| stage, mic ||| best quality, rating: general` とキャプションを指定すると、二番目の `|||` で分割された部分はシャッフル、drop されず末尾に残ります。
   - 既存の機能 `caption_prefix` と `caption_suffix` とあわせて使えます。`caption_prefix` と `caption_suffix` は一番最初に処理され、その後、ワイルドカード、`keep_tokens_separator`、シャッフルおよび drop、`secondary_separator` の順に処理されます。
+  - 詳細は [データセット設定](./docs/config_README-ja.md) をご覧ください。
 - `tag_image_by_wd14_tagger.py` で v3 のリポジトリがサポートされました（`--onnx` 指定時のみ有効）。 PR [#1192](https://github.com/kohya-ss/sd-scripts/pull/1192) sdbds 氏に感謝します。
   - Onnx のバージョンアップが必要になるかもしれません。デフォルトでは Onnx はインストールされていませんので、`pip install onnx==1.15.0 onnxruntime-gpu==1.17.1` 等でインストール、アップデートしてください。`requirements.txt` のコメントもあわせてご確認ください。
 - `tag_image_by_wd14_tagger.py` で、モデルを`--repo_id` のサブディレクトリに保存するようにしました。これにより複数のモデルファイルがキャッシュされます。`--model_dir` 直下の不要なファイルは削除願います。
@@ -286,91 +287,6 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
 - 各学習スクリプトに、学習終了時に state を保存する `--save_state_on_train_end` オプションが追加されました。 PR [#1168](https://github.com/kohya-ss/sd-scripts/pull/1168) gesen2egee 氏に感謝します。
 
 
-#### Example of dataset settings / データセット設定の記述例:
-
-```toml
-[general]
-flip_aug = true
-color_aug = false
-resolution = [1024, 1024]
-
-[[datasets]]
-batch_size = 6
-enable_bucket = true
-bucket_no_upscale = true
-caption_extension = ".txt"
-keep_tokens_separator= "|||"
-shuffle_caption = true
-caption_tag_dropout_rate = 0.1
-secondary_separator = ";;;" # subset 側に書くこともできます / can be written in the subset side
-enable_wildcard = true # 同上 / same as above
-
-  [[datasets.subsets]]
-  image_dir = "/path/to/image_dir"
-  num_repeats = 1
-
-  # ||| の前後はカンマは不要です（自動的に追加されます） / No comma is required before and after ||| (it is added automatically)
-  caption_prefix = "1girl, hatsune miku, vocaloid |||" 
-  
-  # ||| の後はシャッフル、drop されず残ります / After |||, it is not shuffled or dropped and remains
-  # 単純に文字列として連結されるので、カンマなどは自分で入れる必要があります / It is simply concatenated as a string, so you need to put commas yourself
-  caption_suffix = ", anime screencap ||| masterpiece, rating: general"
-```
-
-#### Example of caption, secondary_separator notation: `secondary_separator = ";;;"`
-
-```txt
-1girl, hatsune miku, vocaloid, upper body, looking at viewer, sky;;;cloud;;;day, outdoors
-```
-The part `sky;;;cloud;;;day` is replaced with `sky,cloud,day` without shuffling or dropping. When shuffling and dropping are enabled, it is processed as a whole (as one tag). For example, it becomes `vocaloid, 1girl, upper body, sky,cloud,day, outdoors, hatsune miku` (shuffled) or `vocaloid, 1girl, outdoors, looking at viewer, upper body, hatsune miku` (dropped).
-
-#### Example of caption, enable_wildcard notation: `enable_wildcard = true`
-
-```txt
-1girl, hatsune miku, vocaloid, upper body, looking at viewer, {simple|white} background
-```
-`simple` or `white` is randomly selected, and it becomes `simple background` or `white background`.
-
-```txt
-1girl, hatsune miku, vocaloid, {{retro style}}
-```
-If you want to include `{` or `}` in the tag string, double them like `{{` or `}}` (in this example, the actual caption used for training is `{retro style}`).
-
-#### Example of caption, `keep_tokens_separator` notation: `keep_tokens_separator = "|||"`
-
-```txt
-1girl, hatsune miku, vocaloid ||| stage, microphone, white shirt, smile ||| best quality, rating: general
-```
-It becomes `1girl, hatsune miku, vocaloid, microphone, stage, white shirt, best quality, rating: general` or `1girl, hatsune miku, vocaloid, white shirt, smile, stage, microphone, best quality, rating: general` etc.
-
-
-#### キャプション記述例、secondary_separator 記法：`secondary_separator = ";;;"` の場合
-
-```txt
-1girl, hatsune miku, vocaloid, upper body, looking at viewer, sky;;;cloud;;;day, outdoors
-```
-`sky;;;cloud;;;day` の部分はシャッフル、drop されず `sky,cloud,day` に置換されます。シャッフル、drop が有効な場合、まとめて（一つのタグとして）処理されます。つまり `vocaloid, 1girl, upper body, sky,cloud,day, outdoors, hatsune miku` （シャッフル）や `vocaloid, 1girl, outdoors, looking at viewer, upper body, hatsune miku` （drop されたケース）などになります。
-
-#### キャプション記述例、ワイルドカード記法： `enable_wildcard = true` の場合
-
-```txt
-1girl, hatsune miku, vocaloid, upper body, looking at viewer, {simple|white} background
-```
-ランダムに `simple` または `white` が選ばれ、`simple background` または `white background` になります。
-
-```txt
-1girl, hatsune miku, vocaloid, {{retro style}}
-```
-タグ文字列に `{` や `}` そのものを含めたい場合は `{{` や `}}` のように二つ重ねてください（この例では実際に学習に用いられるキャプションは `{retro style}` になります）。
-
-#### キャプション記述例、`keep_tokens_separator` 記法： `keep_tokens_separator = "|||"` の場合
-
-```txt
-1girl, hatsune miku, vocaloid ||| stage, microphone, white shirt, smile ||| best quality, rating: general
-```
-`1girl, hatsune miku, vocaloid, microphone, stage, white shirt, best quality, rating: general` や `1girl, hatsune miku, vocaloid, white shirt, smile, stage, microphone, best quality, rating: general` などになります。
-
-
 ### Mar 15, 2024 / 2024/3/15: v0.8.5
 
 - Fixed a bug that the value of timestep embedding during SDXL training was incorrect.
diff --git a/docs/config_README-en.md b/docs/config_README-en.md
index a0727934d..bdcaabfc7 100644
--- a/docs/config_README-en.md
+++ b/docs/config_README-en.md
@@ -1,7 +1,10 @@
 Original Source by kohya-ss
 
+First version:
 A.I Translation by Model: NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO, editing by Darkstorm2150
 
+Some parts are manually added.
+
 # Config Readme
 
 This README is about the configuration files that can be passed with the `--dataset_config` option.
@@ -143,11 +146,23 @@ These options are related to subset configuration.
 | `shuffle_caption` | `true` | o | o | o |
 | `caption_prefix` | `"masterpiece, best quality, "` | o | o | o |
 | `caption_suffix` | `", from side"` | o | o | o |
+| `caption_separator` |  (not specified) | o | o | o |
+| `keep_tokens_separator` | `“|||”` | o | o | o |
+| `secondary_separator` | `“;;;”` | o | o | o |
+| `enable_wildcard` | `true` | o | o | o |
 
 * `num_repeats`
     * Specifies the number of repeats for images in a subset. This is equivalent to `--dataset_repeats` in fine-tuning but can be specified for any training method.
 * `caption_prefix`, `caption_suffix`
     * Specifies the prefix and suffix strings to be appended to the captions. Shuffling is performed with these strings included. Be cautious when using `keep_tokens`.
+* `caption_separator`
+    * Specifies the string to separate the tags. The default is `,`. This option is usually not necessary to set.
+* `keep_tokens_separator`
+    * Specifies the string to separate the parts to be fixed in the caption. For example, if you specify `aaa, bbb ||| ccc, ddd, eee, fff ||| ggg, hhh`, the parts `aaa, bbb` and `ggg, hhh` will remain, and the rest will be shuffled and dropped. The comma in between is not necessary. As a result, the prompt will be `aaa, bbb, eee, ccc, fff, ggg, hhh` or `aaa, bbb, fff, ccc, eee, ggg, hhh`, etc.
+* `secondary_separator`
+    * Specifies an additional separator. The part separated by this separator is treated as one tag and is shuffled and dropped. It is then replaced by `caption_separator`. For example, if you specify `aaa;;;bbb;;;ccc`, it will be replaced by `aaa,bbb,ccc` or dropped together.
+* `enable_wildcard`
+    * Enables wildcard notation. This will be explained later.
 
 ### DreamBooth-specific options
 
@@ -276,4 +291,62 @@ As a temporary measure, we will list common errors and their solutions. If you e
 * `voluptuous.error.MultipleInvalid: expected int for dictionary value @ ...`: This error occurs when the specified value format is incorrect. It is highly likely that the value format is incorrect. The `int` part changes depending on the target option. The example configurations in this README may be helpful.
 * `voluptuous.error.MultipleInvalid: extra keys not allowed @ ...`: This error occurs when there is an option name that is not supported. It is highly likely that you misspelled the option name or mistakenly included it.
 
+## Miscellaneous
+
+### Example of configuration file, 設定ファイルの記述例
+
+```toml
+[general]
+flip_aug = true
+color_aug = false
+resolution = [1024, 1024]
+
+[[datasets]]
+batch_size = 6
+enable_bucket = true
+bucket_no_upscale = true
+caption_extension = ".txt"
+keep_tokens_separator= "|||"
+shuffle_caption = true
+caption_tag_dropout_rate = 0.1
+secondary_separator = ";;;" # subset 側に書くこともできます / can be written in the subset side
+enable_wildcard = true # 同上 / same as above
+
+  [[datasets.subsets]]
+  image_dir = "/path/to/image_dir"
+  num_repeats = 1
+
+  # ||| の前後はカンマは不要です（自動的に追加されます） / No comma is required before and after ||| (it is added automatically)
+  caption_prefix = "1girl, hatsune miku, vocaloid |||" 
+  
+  # ||| の後はシャッフル、drop されず残ります / After |||, it is not shuffled or dropped and remains
+  # 単純に文字列として連結されるので、カンマなどは自分で入れる必要があります / It is simply concatenated as a string, so you need to put commas yourself
+  caption_suffix = ", anime screencap ||| masterpiece, rating: general"
+```
+
+### Example of caption, secondary_separator notation: `secondary_separator = ";;;"`
+
+```txt
+1girl, hatsune miku, vocaloid, upper body, looking at viewer, sky;;;cloud;;;day, outdoors
+```
+The part `sky;;;cloud;;;day` is replaced with `sky,cloud,day` without shuffling or dropping. When shuffling and dropping are enabled, it is processed as a whole (as one tag). For example, it becomes `vocaloid, 1girl, upper body, sky,cloud,day, outdoors, hatsune miku` (shuffled) or `vocaloid, 1girl, outdoors, looking at viewer, upper body, hatsune miku` (dropped).
+
+### Example of caption, enable_wildcard notation: `enable_wildcard = true`
+
+```txt
+1girl, hatsune miku, vocaloid, upper body, looking at viewer, {simple|white} background
+```
+`simple` or `white` is randomly selected, and it becomes `simple background` or `white background`.
+
+```txt
+1girl, hatsune miku, vocaloid, {{retro style}}
+```
+If you want to include `{` or `}` in the tag string, double them like `{{` or `}}` (in this example, the actual caption used for training is `{retro style}`).
+
+### Example of caption, `keep_tokens_separator` notation: `keep_tokens_separator = "|||"`
+
+```txt
+1girl, hatsune miku, vocaloid ||| stage, microphone, white shirt, smile ||| best quality, rating: general
+```
+It becomes `1girl, hatsune miku, vocaloid, microphone, stage, white shirt, best quality, rating: general` or `1girl, hatsune miku, vocaloid, white shirt, smile, stage, microphone, best quality, rating: general` etc.
 
diff --git a/docs/config_README-ja.md b/docs/config_README-ja.md
index 69a03f6cf..47bb5c57d 100644
--- a/docs/config_README-ja.md
+++ b/docs/config_README-ja.md
@@ -1,5 +1,3 @@
-For non-Japanese speakers: this README is provided only in Japanese in the current state. Sorry for inconvenience. We will provide English version in the near future.
-
 `--dataset_config` で渡すことができる設定ファイルに関する説明です。
 
 ## 概要
@@ -140,12 +138,28 @@ DreamBooth の手法と fine tuning の手法の両方とも利用可能な学
 | `shuffle_caption` | `true` | o | o | o |
 | `caption_prefix` | `“masterpiece, best quality, ”` | o | o | o |
 | `caption_suffix` | `“, from side”` | o | o | o |
+| `caption_separator` | （通常は設定しません） | o | o | o |
+| `keep_tokens_separator` | `“|||”` | o | o | o |
+| `secondary_separator` | `“;;;”` | o | o | o |
+| `enable_wildcard` | `true` | o | o | o |
 
 * `num_repeats`
     * サブセットの画像の繰り返し回数を指定します。fine tuning における `--dataset_repeats` に相当しますが、`num_repeats` はどの学習方法でも指定可能です。
 * `caption_prefix`, `caption_suffix`
     * キャプションの前、後に付与する文字列を指定します。シャッフルはこれらの文字列を含めた状態で行われます。`keep_tokens` を指定する場合には注意してください。
 
+* `caption_separator`
+    * タグを区切る文字列を指定します。デフォルトは `,` です。このオプションは通常は設定する必要はありません。
+
+* `keep_tokens_separator`
+    *  キャプションで固定したい部分を区切る文字列を指定します。たとえば `aaa, bbb ||| ccc, ddd, eee, fff ||| ggg, hhh` のように指定すると、`aaa, bbb` と `ggg, hhh` の部分はシャッフル、drop されず残ります。間のカンマは不要です。結果としてプロンプトは `aaa, bbb, eee, ccc, fff, ggg, hhh` や `aaa, bbb, fff, ccc, eee, ggg, hhh` などになります。
+
+* `secondary_separator`
+    * 追加の区切り文字を指定します。この区切り文字で区切られた部分は一つのタグとして扱われ、シャッフル、drop されます。その後、`caption_separator` に置き換えられます。たとえば `aaa;;;bbb;;;ccc` のように指定すると、`aaa,bbb,ccc` に置き換えられるか、まとめて drop されます。
+
+* `enable_wildcard`
+    * ワイルドカード記法を有効にします。ワイルドカード記法については後述します。
+
 ### DreamBooth 方式専用のオプション
 
 DreamBooth 方式のオプションは、サブセット向けオプションのみ存在します。
@@ -280,4 +294,61 @@ resolution = 768
 * `voluptuous.error.MultipleInvalid: expected int for dictionary value @ ...`: 指定する値の形式が不正というエラーです。値の形式が間違っている可能性が高いです。`int` の部分は対象となるオプションによって変わります。この README に載っているオプションの「設定例」が役立つかもしれません。
 * `voluptuous.error.MultipleInvalid: extra keys not allowed @ ...`: 対応していないオプション名が存在している場合に発生するエラーです。オプション名を間違って記述しているか、誤って紛れ込んでいる可能性が高いです。
 
+## その他
+
+### Example of configuration file, 設定ファイルの記述例
 
+```toml
+[general]
+flip_aug = true
+color_aug = false
+resolution = [1024, 1024]
+
+[[datasets]]
+batch_size = 6
+enable_bucket = true
+bucket_no_upscale = true
+caption_extension = ".txt"
+keep_tokens_separator= "|||"
+shuffle_caption = true
+caption_tag_dropout_rate = 0.1
+secondary_separator = ";;;" # subset 側に書くこともできます / can be written in the subset side
+enable_wildcard = true # 同上 / same as above
+
+  [[datasets.subsets]]
+  image_dir = "/path/to/image_dir"
+  num_repeats = 1
+
+  # ||| の前後はカンマは不要です（自動的に追加されます） / No comma is required before and after ||| (it is added automatically)
+  caption_prefix = "1girl, hatsune miku, vocaloid |||" 
+  
+  # ||| の後はシャッフル、drop されず残ります / After |||, it is not shuffled or dropped and remains
+  # 単純に文字列として連結されるので、カンマなどは自分で入れる必要があります / It is simply concatenated as a string, so you need to put commas yourself
+  caption_suffix = ", anime screencap ||| masterpiece, rating: general"
+```
+
+### キャプション記述例、secondary_separator 記法：`secondary_separator = ";;;"` の場合
+
+```txt
+1girl, hatsune miku, vocaloid, upper body, looking at viewer, sky;;;cloud;;;day, outdoors
+```
+`sky;;;cloud;;;day` の部分はシャッフル、drop されず `sky,cloud,day` に置換されます。シャッフル、drop が有効な場合、まとめて（一つのタグとして）処理されます。つまり `vocaloid, 1girl, upper body, sky,cloud,day, outdoors, hatsune miku` （シャッフル）や `vocaloid, 1girl, outdoors, looking at viewer, upper body, hatsune miku` （drop されたケース）などになります。
+
+### キャプション記述例、ワイルドカード記法： `enable_wildcard = true` の場合
+
+```txt
+1girl, hatsune miku, vocaloid, upper body, looking at viewer, {simple|white} background
+```
+ランダムに `simple` または `white` が選ばれ、`simple background` または `white background` になります。
+
+```txt
+1girl, hatsune miku, vocaloid, {{retro style}}
+```
+タグ文字列に `{` や `}` そのものを含めたい場合は `{{` や `}}` のように二つ重ねてください（この例では実際に学習に用いられるキャプションは `{retro style}` になります）。
+
+### キャプション記述例、`keep_tokens_separator` 記法： `keep_tokens_separator = "|||"` の場合
+
+```txt
+1girl, hatsune miku, vocaloid ||| stage, microphone, white shirt, smile ||| best quality, rating: general
+```
+`1girl, hatsune miku, vocaloid, microphone, stage, white shirt, best quality, rating: general` や `1girl, hatsune miku, vocaloid, white shirt, smile, stage, microphone, best quality, rating: general` などになります。

From 594c7f70500e402586654e73501e7d8fc74592b8 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sat, 23 Mar 2024 16:11:31 +0900
Subject: [PATCH 39/69] format by black

---
 finetune/merge_captions_to_metadata.py | 132 ++++++++++++++-----------
 finetune/merge_dd_tags_to_metadata.py  | 118 ++++++++++++----------
 2 files changed, 144 insertions(+), 106 deletions(-)

diff --git a/finetune/merge_captions_to_metadata.py b/finetune/merge_captions_to_metadata.py
index 60765b863..89f717473 100644
--- a/finetune/merge_captions_to_metadata.py
+++ b/finetune/merge_captions_to_metadata.py
@@ -6,75 +6,95 @@
 import library.train_util as train_util
 import os
 from library.utils import setup_logging
+
 setup_logging()
 import logging
+
 logger = logging.getLogger(__name__)
 
+
 def main(args):
-  assert not args.recursive or (args.recursive and args.full_path), "recursive requires full_path / recursiveはfull_pathと同時に指定してください"
+    assert not args.recursive or (
+        args.recursive and args.full_path
+    ), "recursive requires full_path / recursiveはfull_pathと同時に指定してください"
 
-  train_data_dir_path = Path(args.train_data_dir)
-  image_paths: List[Path] = train_util.glob_images_pathlib(train_data_dir_path, args.recursive)
-  logger.info(f"found {len(image_paths)} images.")
+    train_data_dir_path = Path(args.train_data_dir)
+    image_paths: List[Path] = train_util.glob_images_pathlib(train_data_dir_path, args.recursive)
+    logger.info(f"found {len(image_paths)} images.")
 
-  if args.in_json is None and Path(args.out_json).is_file():
-    args.in_json = args.out_json
+    if args.in_json is None and Path(args.out_json).is_file():
+        args.in_json = args.out_json
 
-  if args.in_json is not None:
-    logger.info(f"loading existing metadata: {args.in_json}")
-    metadata = json.loads(Path(args.in_json).read_text(encoding='utf-8'))
-    logger.warning("captions for existing images will be overwritten / 既存の画像のキャプションは上書きされます")
-  else:
-    logger.info("new metadata will be created / 新しいメタデータファイルが作成されます")
-    metadata = {}
+    if args.in_json is not None:
+        logger.info(f"loading existing metadata: {args.in_json}")
+        metadata = json.loads(Path(args.in_json).read_text(encoding="utf-8"))
+        logger.warning("captions for existing images will be overwritten / 既存の画像のキャプションは上書きされます")
+    else:
+        logger.info("new metadata will be created / 新しいメタデータファイルが作成されます")
+        metadata = {}
 
-  logger.info("merge caption texts to metadata json.")
-  for image_path in tqdm(image_paths):
-    caption_path = image_path.with_suffix(args.caption_extension)
-    caption = caption_path.read_text(encoding='utf-8').strip()
+    logger.info("merge caption texts to metadata json.")
+    for image_path in tqdm(image_paths):
+        caption_path = image_path.with_suffix(args.caption_extension)
+        caption = caption_path.read_text(encoding="utf-8").strip()
 
-    if not os.path.exists(caption_path):
-      caption_path = os.path.join(image_path, args.caption_extension)
+        if not os.path.exists(caption_path):
+            caption_path = os.path.join(image_path, args.caption_extension)
 
-    image_key = str(image_path) if args.full_path else image_path.stem
-    if image_key not in metadata:
-      metadata[image_key] = {}
+        image_key = str(image_path) if args.full_path else image_path.stem
+        if image_key not in metadata:
+            metadata[image_key] = {}
 
-    metadata[image_key]['caption'] = caption
-    if args.debug:
-      logger.info(f"{image_key} {caption}")
+        metadata[image_key]["caption"] = caption
+        if args.debug:
+            logger.info(f"{image_key} {caption}")
 
-  # metadataを書き出して終わり
-  logger.info(f"writing metadata: {args.out_json}")
-  Path(args.out_json).write_text(json.dumps(metadata, indent=2), encoding='utf-8')
-  logger.info("done!")
+    # metadataを書き出して終わり
+    logger.info(f"writing metadata: {args.out_json}")
+    Path(args.out_json).write_text(json.dumps(metadata, indent=2), encoding="utf-8")
+    logger.info("done!")
 
 
 def setup_parser() -> argparse.ArgumentParser:
-  parser = argparse.ArgumentParser()
-  parser.add_argument("train_data_dir", type=str, help="directory for train images / 学習画像データのディレクトリ")
-  parser.add_argument("out_json", type=str, help="metadata file to output / メタデータファイル書き出し先")
-  parser.add_argument("--in_json", type=str,
-                      help="metadata file to input (if omitted and out_json exists, existing out_json is read) / 読み込むメタデータファイル（省略時、out_jsonが存在すればそれを読み込む）")
-  parser.add_argument("--caption_extention", type=str, default=None,
-                      help="extension of caption file (for backward compatibility) / 読み込むキャプションファイルの拡張子（スペルミスしていたのを残してあります）")
-  parser.add_argument("--caption_extension", type=str, default=".caption", help="extension of caption file / 読み込むキャプションファイルの拡張子")
-  parser.add_argument("--full_path", action="store_true",
-                      help="use full path as image-key in metadata (supports multiple directories) / メタデータで画像キーをフルパスにする（複数の学習画像ディレクトリに対応）")
-  parser.add_argument("--recursive", action="store_true",
-                      help="recursively look for training tags in all child folders of train_data_dir / train_data_dirのすべての子フォルダにある学習タグを再帰的に探す")
-  parser.add_argument("--debug", action="store_true", help="debug mode")
-
-  return parser
-
-
-if __name__ == '__main__':
-  parser = setup_parser()
-
-  args = parser.parse_args()
-
-  # スペルミスしていたオプションを復元する
-  if args.caption_extention is not None:
-    args.caption_extension = args.caption_extention
-
-  main(args)
+    parser = argparse.ArgumentParser()
+    parser.add_argument("train_data_dir", type=str, help="directory for train images / 学習画像データのディレクトリ")
+    parser.add_argument("out_json", type=str, help="metadata file to output / メタデータファイル書き出し先")
+    parser.add_argument(
+        "--in_json",
+        type=str,
+        help="metadata file to input (if omitted and out_json exists, existing out_json is read) / 読み込むメタデータファイル（省略時、out_jsonが存在すればそれを読み込む）",
+    )
+    parser.add_argument(
+        "--caption_extention",
+        type=str,
+        default=None,
+        help="extension of caption file (for backward compatibility) / 読み込むキャプションファイルの拡張子（スペルミスしていたのを残してあります）",
+    )
+    parser.add_argument(
+        "--caption_extension", type=str, default=".caption", help="extension of caption file / 読み込むキャプションファイルの拡張子"
+    )
+    parser.add_argument(
+        "--full_path",
+        action="store_true",
+        help="use full path as image-key in metadata (supports multiple directories) / メタデータで画像キーをフルパスにする（複数の学習画像ディレクトリに対応）",
+    )
+    parser.add_argument(
+        "--recursive",
+        action="store_true",
+        help="recursively look for training tags in all child folders of train_data_dir / train_data_dirのすべての子フォルダにある学習タグを再帰的に探す",
+    )
+    parser.add_argument("--debug", action="store_true", help="debug mode")
+
+    return parser
+
+
+if __name__ == "__main__":
+    parser = setup_parser()
+
+    args = parser.parse_args()
+
+    # スペルミスしていたオプションを復元する
+    if args.caption_extention is not None:
+        args.caption_extension = args.caption_extention
+
+    main(args)
diff --git a/finetune/merge_dd_tags_to_metadata.py b/finetune/merge_dd_tags_to_metadata.py
index 9ef8f14b0..ce22d990e 100644
--- a/finetune/merge_dd_tags_to_metadata.py
+++ b/finetune/merge_dd_tags_to_metadata.py
@@ -6,70 +6,88 @@
 import library.train_util as train_util
 import os
 from library.utils import setup_logging
+
 setup_logging()
 import logging
+
 logger = logging.getLogger(__name__)
 
+
 def main(args):
-  assert not args.recursive or (args.recursive and args.full_path), "recursive requires full_path / recursiveはfull_pathと同時に指定してください"
+    assert not args.recursive or (
+        args.recursive and args.full_path
+    ), "recursive requires full_path / recursiveはfull_pathと同時に指定してください"
 
-  train_data_dir_path = Path(args.train_data_dir)
-  image_paths: List[Path] = train_util.glob_images_pathlib(train_data_dir_path, args.recursive)
-  logger.info(f"found {len(image_paths)} images.")
+    train_data_dir_path = Path(args.train_data_dir)
+    image_paths: List[Path] = train_util.glob_images_pathlib(train_data_dir_path, args.recursive)
+    logger.info(f"found {len(image_paths)} images.")
 
-  if args.in_json is None and Path(args.out_json).is_file():
-    args.in_json = args.out_json
+    if args.in_json is None and Path(args.out_json).is_file():
+        args.in_json = args.out_json
 
-  if args.in_json is not None:
-    logger.info(f"loading existing metadata: {args.in_json}")
-    metadata = json.loads(Path(args.in_json).read_text(encoding='utf-8'))
-    logger.warning("tags data for existing images will be overwritten / 既存の画像のタグは上書きされます")
-  else:
-    logger.info("new metadata will be created / 新しいメタデータファイルが作成されます")
-    metadata = {}
+    if args.in_json is not None:
+        logger.info(f"loading existing metadata: {args.in_json}")
+        metadata = json.loads(Path(args.in_json).read_text(encoding="utf-8"))
+        logger.warning("tags data for existing images will be overwritten / 既存の画像のタグは上書きされます")
+    else:
+        logger.info("new metadata will be created / 新しいメタデータファイルが作成されます")
+        metadata = {}
 
-  logger.info("merge tags to metadata json.")
-  for image_path in tqdm(image_paths):
-    tags_path = image_path.with_suffix(args.caption_extension)
-    tags = tags_path.read_text(encoding='utf-8').strip()
+    logger.info("merge tags to metadata json.")
+    for image_path in tqdm(image_paths):
+        tags_path = image_path.with_suffix(args.caption_extension)
+        tags = tags_path.read_text(encoding="utf-8").strip()
 
-    if not os.path.exists(tags_path):
-      tags_path = os.path.join(image_path, args.caption_extension)
+        if not os.path.exists(tags_path):
+            tags_path = os.path.join(image_path, args.caption_extension)
 
-    image_key = str(image_path) if args.full_path else image_path.stem
-    if image_key not in metadata:
-      metadata[image_key] = {}
+        image_key = str(image_path) if args.full_path else image_path.stem
+        if image_key not in metadata:
+            metadata[image_key] = {}
 
-    metadata[image_key]['tags'] = tags
-    if args.debug:
-      logger.info(f"{image_key} {tags}")
+        metadata[image_key]["tags"] = tags
+        if args.debug:
+            logger.info(f"{image_key} {tags}")
 
-  # metadataを書き出して終わり
-  logger.info(f"writing metadata: {args.out_json}")
-  Path(args.out_json).write_text(json.dumps(metadata, indent=2), encoding='utf-8')
+    # metadataを書き出して終わり
+    logger.info(f"writing metadata: {args.out_json}")
+    Path(args.out_json).write_text(json.dumps(metadata, indent=2), encoding="utf-8")
 
-  logger.info("done!")
+    logger.info("done!")
 
 
 def setup_parser() -> argparse.ArgumentParser:
-  parser = argparse.ArgumentParser()
-  parser.add_argument("train_data_dir", type=str, help="directory for train images / 学習画像データのディレクトリ")
-  parser.add_argument("out_json", type=str, help="metadata file to output / メタデータファイル書き出し先")
-  parser.add_argument("--in_json", type=str,
-                      help="metadata file to input (if omitted and out_json exists, existing out_json is read) / 読み込むメタデータファイル（省略時、out_jsonが存在すればそれを読み込む）")
-  parser.add_argument("--full_path", action="store_true",
-                      help="use full path as image-key in metadata (supports multiple directories) / メタデータで画像キーをフルパスにする（複数の学習画像ディレクトリに対応）")
-  parser.add_argument("--recursive", action="store_true",
-                      help="recursively look for training tags in all child folders of train_data_dir / train_data_dirのすべての子フォルダにある学習タグを再帰的に探す")
-  parser.add_argument("--caption_extension", type=str, default=".txt",
-                      help="extension of caption (tag) file / 読み込むキャプション（タグ）ファイルの拡張子")
-  parser.add_argument("--debug", action="store_true", help="debug mode, print tags")
-
-  return parser
-
-
-if __name__ == '__main__':
-  parser = setup_parser()
-
-  args = parser.parse_args()
-  main(args)
+    parser = argparse.ArgumentParser()
+    parser.add_argument("train_data_dir", type=str, help="directory for train images / 学習画像データのディレクトリ")
+    parser.add_argument("out_json", type=str, help="metadata file to output / メタデータファイル書き出し先")
+    parser.add_argument(
+        "--in_json",
+        type=str,
+        help="metadata file to input (if omitted and out_json exists, existing out_json is read) / 読み込むメタデータファイル（省略時、out_jsonが存在すればそれを読み込む）",
+    )
+    parser.add_argument(
+        "--full_path",
+        action="store_true",
+        help="use full path as image-key in metadata (supports multiple directories) / メタデータで画像キーをフルパスにする（複数の学習画像ディレクトリに対応）",
+    )
+    parser.add_argument(
+        "--recursive",
+        action="store_true",
+        help="recursively look for training tags in all child folders of train_data_dir / train_data_dirのすべての子フォルダにある学習タグを再帰的に探す",
+    )
+    parser.add_argument(
+        "--caption_extension",
+        type=str,
+        default=".txt",
+        help="extension of caption (tag) file / 読み込むキャプション（タグ）ファイルの拡張子",
+    )
+    parser.add_argument("--debug", action="store_true", help="debug mode, print tags")
+
+    return parser
+
+
+if __name__ == "__main__":
+    parser = setup_parser()
+
+    args = parser.parse_args()
+    main(args)

From f4a4c11cd30a885d1d5ddb86bee609305c5398f3 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sat, 23 Mar 2024 18:51:37 +0900
Subject: [PATCH 40/69] support multiline captions ref #1155

---
 README.md                |  8 ++++----
 docs/config_README-en.md | 30 +++++++++++++++++++++++++++++-
 docs/config_README-ja.md | 32 ++++++++++++++++++++++++++++++--
 library/train_util.py    | 38 +++++++++++++++++++++++++++++++-------
 4 files changed, 94 insertions(+), 14 deletions(-)

diff --git a/README.md b/README.md
index dae311325..dd000d126 100644
--- a/README.md
+++ b/README.md
@@ -257,8 +257,8 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
 - `train_network.py` and `sdxl_train_network.py` are modified to record some dataset settings in the metadata of the trained model (`caption_prefix`, `caption_suffix`, `keep_tokens_separator`, `secondary_separator`, `enable_wildcard`).
 - Some features are added to the dataset subset settings.
   - `secondary_separator` is added to specify the tag separator that is not the target of shuffling or dropping. 
-    - Specify `secondary_separator=";;;"`. When you specify `secondary_separator`, the part is not shuffled or dropped. See the example below.
-  - `enable_wildcard` is added. When set to `true`, the wildcard notation `{aaa|bbb|ccc}` can be used. See the example below.
+    - Specify `secondary_separator=";;;"`. When you specify `secondary_separator`, the part is not shuffled or dropped. 
+  - `enable_wildcard` is added. When set to `true`, the wildcard notation `{aaa|bbb|ccc}` can be used. The multi-line caption is also enabled.
   - `keep_tokens_separator` is updated to be used twice in the caption. When you specify `keep_tokens_separator="|||"`, the part divided by the second `|||` is not shuffled or dropped and remains at the end.
   - The existing features `caption_prefix` and `caption_suffix` can be used together. `caption_prefix` and `caption_suffix` are processed first, and then `enable_wildcard`, `keep_tokens_separator`, shuffling and dropping, and `secondary_separator` are processed in order.
   - See [Dataset config](./docs/config_README-en.md) for details.
@@ -274,8 +274,8 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
 - データセット設定の `.toml` ファイルが UTF-8 encoding で読み込まれるようになりました。PR [#1167](https://github.com/kohya-ss/sd-scripts/pull/1167) Horizon1704 氏に感謝します。
 - `train_network.py` および `sdxl_train_network.py` で、学習したモデルのメタデータに一部のデータセット設定が記録されるよう修正しました（`caption_prefix`、`caption_suffix`、`keep_tokens_separator`、`secondary_separator`、`enable_wildcard`）。
 - データセットのサブセット設定にいくつかの機能を追加しました。
-  - シャッフルの対象とならないタグ分割識別子の指定 `secondary_separator` を追加しました。`secondary_separator=";;;"` のように指定します。`secondary_separator` で区切ることで、その部分はシャッフル、drop 時にまとめて扱われます。詳しくは記述例をご覧ください。
-  - `enable_wildcard` を追加しました。`true` にするとワイルドカード記法 `{aaa|bbb|ccc}` が使えます。詳しくは記述例をご覧ください。
+  - シャッフルの対象とならないタグ分割識別子の指定 `secondary_separator` を追加しました。`secondary_separator=";;;"` のように指定します。`secondary_separator` で区切ることで、その部分はシャッフル、drop 時にまとめて扱われます。
+  - `enable_wildcard` を追加しました。`true` にするとワイルドカード記法 `{aaa|bbb|ccc}` が使えます。また複数行キャプションも有効になります。
   - `keep_tokens_separator` をキャプション内に 2 つ使えるようにしました。たとえば `keep_tokens_separator="|||"` と指定したとき、`1girl, hatsune miku, vocaloid ||| stage, mic ||| best quality, rating: general` とキャプションを指定すると、二番目の `|||` で分割された部分はシャッフル、drop されず末尾に残ります。
   - 既存の機能 `caption_prefix` と `caption_suffix` とあわせて使えます。`caption_prefix` と `caption_suffix` は一番最初に処理され、その後、ワイルドカード、`keep_tokens_separator`、シャッフルおよび drop、`secondary_separator` の順に処理されます。
   - 詳細は [データセット設定](./docs/config_README-ja.md) をご覧ください。
diff --git a/docs/config_README-en.md b/docs/config_README-en.md
index bdcaabfc7..e99fde216 100644
--- a/docs/config_README-en.md
+++ b/docs/config_README-en.md
@@ -293,7 +293,35 @@ As a temporary measure, we will list common errors and their solutions. If you e
 
 ## Miscellaneous
 
-### Example of configuration file, 設定ファイルの記述例
+### Multi-line captions
+
+By setting `enable_wildcard = true`, multiple-line captions are also enabled. If the caption file consists of multiple lines, one line is randomly selected as the caption. 
+
+```txt
+1girl, hatsune miku, vocaloid, upper body, looking at viewer, microphone, stage
+a girl with a microphone standing on a stage
+detailed digital art of a girl with a microphone on a stage
+```
+
+It can be combined with wildcard notation.
+
+In metadata files, you can also specify multiple-line captions. In the `.json` metadata file, use `\n` to represent a line break. If the caption file consists of multiple lines, `merge_captions_to_metadata.py` will create a metadata file in this format.
+
+The tags in the metadata (`tags`) are added to each line of the caption.
+
+```json
+{
+    "/path/to/image.png": {
+        "caption": "a cartoon of a frog with the word frog on it\ntest multiline caption1\ntest multiline caption2",
+        "tags": "open mouth, simple background, standing, no humans, animal, black background, frog, animal costume, animal focus"
+    },
+    ...
+}
+```
+
+In this case, the actual caption will be `a cartoon of a frog with the word frog on it, open mouth, simple background ...`, `test multiline caption1, open mouth, simple background ...`, `test multiline caption2, open mouth, simple background ...`, etc.
+
+### Example of configuration file : `secondary_separator`, wildcard notation, `keep_tokens_separator`, etc.
 
 ```toml
 [general]
diff --git a/docs/config_README-ja.md b/docs/config_README-ja.md
index 47bb5c57d..b57ae86a7 100644
--- a/docs/config_README-ja.md
+++ b/docs/config_README-ja.md
@@ -158,7 +158,7 @@ DreamBooth の手法と fine tuning の手法の両方とも利用可能な学
     * 追加の区切り文字を指定します。この区切り文字で区切られた部分は一つのタグとして扱われ、シャッフル、drop されます。その後、`caption_separator` に置き換えられます。たとえば `aaa;;;bbb;;;ccc` のように指定すると、`aaa,bbb,ccc` に置き換えられるか、まとめて drop されます。
 
 * `enable_wildcard`
-    * ワイルドカード記法を有効にします。ワイルドカード記法については後述します。
+    * ワイルドカード記法および複数行キャプションを有効にします。ワイルドカード記法、複数行キャプションについては後述します。
 
 ### DreamBooth 方式専用のオプション
 
@@ -296,7 +296,35 @@ resolution = 768
 
 ## その他
 
-### Example of configuration file, 設定ファイルの記述例
+### 複数行キャプション
+
+`enable_wildcard = true` を設定することで、複数行キャプションも同時に有効になります。キャプションファイルが複数の行からなる場合、ランダムに一つの行が選ばれてキャプションとして利用されます。
+
+```txt
+1girl, hatsune miku, vocaloid, upper body, looking at viewer, microphone, stage
+a girl with a microphone standing on a stage
+detailed digital art of a girl with a microphone on a stage
+```
+
+ワイルドカード記法と組み合わせることも可能です。
+
+メタデータファイルでも同様に複数行キャプションを指定することができます。メタデータの .json 内には、`\n` を使って改行を表現してください。キャプションファイルが複数行からなる場合、`merge_captions_to_metadata.py` を使うと、この形式でメタデータファイルが作成されます。
+
+メタデータのタグ (`tags`) は、キャプションの各行に追加されます。
+
+```json
+{
+    "/path/to/image.png": {
+        "caption": "a cartoon of a frog with the word frog on it\ntest multiline caption1\ntest multiline caption2",
+        "tags": "open mouth, simple background, standing, no humans, animal, black background, frog, animal costume, animal focus"
+    },
+    ...
+}
+```
+
+この場合、実際のキャプションは `a cartoon of a frog with the word frog on it, open mouth, simple background ...` または `test multiline caption1, open mouth, simple background ...`、 `test multiline caption2, open mouth, simple background ...` 等になります。
+
+### 設定ファイルの記述例：追加の区切り文字、ワイルドカード記法、`keep_tokens_separator` 等
 
 ```toml
 [general]
diff --git a/library/train_util.py b/library/train_util.py
index a13985ee2..d076cf847 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -693,6 +693,10 @@ def process_caption(self, subset: BaseSubset, caption):
         else:
             # process wildcards
             if subset.enable_wildcard:
+                # if caption is multiline, random choice one line
+                if "\n" in caption:
+                    caption = random.choice(caption.split("\n"))
+
                 # wildcard is like '{aaa|bbb|ccc...}'
                 # escape the curly braces like {{ or }}
                 replacer1 = "⦅"
@@ -711,6 +715,9 @@ def replace_wildcard(match):
 
                 # unescape the curly braces
                 caption = caption.replace(replacer1, "{").replace(replacer2, "}")
+            else:
+                # if caption is multiline, use the first line
+                caption = caption.split("\n")[0]
 
             if subset.shuffle_caption or subset.token_warmup_step > 0 or subset.caption_tag_dropout_rate > 0:
                 fixed_tokens = []
@@ -1446,7 +1453,7 @@ def __init__(
             self.bucket_reso_steps = None  # この情報は使われない
             self.bucket_no_upscale = False
 
-        def read_caption(img_path, caption_extension):
+        def read_caption(img_path, caption_extension, enable_wildcard):
             # captionの候補ファイル名を作る
             base_name = os.path.splitext(img_path)[0]
             base_name_face_det = base_name
@@ -1465,7 +1472,10 @@ def read_caption(img_path, caption_extension):
                             logger.error(f"illegal char in file (not UTF-8) / ファイルにUTF-8以外の文字があります: {cap_path}")
                             raise e
                         assert len(lines) > 0, f"caption file is empty / キャプションファイルが空です: {cap_path}"
-                        caption = lines[0].strip()
+                        if enable_wildcard:
+                            caption = "\n".join([line.strip() for line in lines if line.strip() != ""]) # 空行を除く、改行で連結
+                        else:
+                            caption = lines[0].strip()
                     break
             return caption
 
@@ -1481,7 +1491,7 @@ def load_dreambooth_dir(subset: DreamBoothSubset):
             captions = []
             missing_captions = []
             for img_path in img_paths:
-                cap_for_img = read_caption(img_path, subset.caption_extension)
+                cap_for_img = read_caption(img_path, subset.caption_extension, subset.enable_wildcard)
                 if cap_for_img is None and subset.class_tokens is None:
                     logger.warning(
                         f"neither caption file nor class tokens are found. use empty caption for {img_path} / キャプションファイルもclass tokenも見つかりませんでした。空のキャプションを使用します: {img_path}"
@@ -1657,10 +1667,24 @@ def __init__(
                 caption = img_md.get("caption")
                 tags = img_md.get("tags")
                 if caption is None:
-                    caption = tags
-                elif tags is not None and len(tags) > 0:
-                    caption = caption + ", " + tags
-                    tags_list.append(tags)
+                    caption = tags  # could be multiline
+                    tags = None
+
+                if subset.enable_wildcard:
+                    # tags must be single line
+                    if tags is not None:
+                        tags = tags.replace("\n", subset.caption_separator)
+
+                    # add tags to each line of caption
+                    if caption is not None and tags is not None:
+                        caption = "\n".join(
+                            [f"{line}{subset.caption_separator}{tags}" for line in caption.split("\n") if line.strip() != ""]
+                        )
+                else:
+                    # use as is
+                    if tags is not None and len(tags) > 0:
+                        caption = caption + subset.caption_separator + tags
+                        tags_list.append(tags)
 
                 if caption is None:
                     caption = ""

From 0c7baea88cfa98c5fab2898551c426f2d4fac4c6 Mon Sep 17 00:00:00 2001
From: feffy380 <114889020+feffy380@users.noreply.github.com>
Date: Sat, 23 Mar 2024 17:28:02 +0100
Subject: [PATCH 41/69] register reg images with correct subset

---
 library/train_util.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/library/train_util.py b/library/train_util.py
index d076cf847..b69fb0950 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -1554,7 +1554,7 @@ def load_dreambooth_dir(subset: DreamBoothSubset):
             for img_path, caption in zip(img_paths, captions):
                 info = ImageInfo(img_path, subset.num_repeats, caption, subset.is_reg, img_path)
                 if subset.is_reg:
-                    reg_infos.append(info)
+                    reg_infos.append((info, subset))
                 else:
                     self.register_image(info, subset)
 
@@ -1575,7 +1575,7 @@ def load_dreambooth_dir(subset: DreamBoothSubset):
             n = 0
             first_loop = True
             while n < num_train_images:
-                for info in reg_infos:
+                for info, subset in reg_infos:
                     if first_loop:
                         self.register_image(info, subset)
                         n += info.num_repeats

From 79d1c12ab056e7114257d7079f2f8846e329320e Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sun, 24 Mar 2024 11:06:37 +0900
Subject: [PATCH 42/69] disable sample_every_n_xxx if value less than 1 ref
 #1202

---
 library/train_util.py | 14 +++++++++++++-
 1 file changed, 13 insertions(+), 1 deletion(-)

diff --git a/library/train_util.py b/library/train_util.py
index d076cf847..8fbf3283d 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -1473,7 +1473,7 @@ def read_caption(img_path, caption_extension, enable_wildcard):
                             raise e
                         assert len(lines) > 0, f"caption file is empty / キャプションファイルが空です: {cap_path}"
                         if enable_wildcard:
-                            caption = "\n".join([line.strip() for line in lines if line.strip() != ""]) # 空行を除く、改行で連結
+                            caption = "\n".join([line.strip() for line in lines if line.strip() != ""])  # 空行を除く、改行で連結
                         else:
                             caption = lines[0].strip()
                     break
@@ -3338,6 +3338,18 @@ def verify_training_args(args: argparse.Namespace):
             + " / zero_terminal_snrが有効ですが、v_parameterizationが有効ではありません。学習結果は想定外になる可能性があります"
         )
 
+    if args.sample_every_n_epochs is not None and args.sample_every_n_epochs <= 0:
+        logger.warning(
+            "sample_every_n_epochs is less than or equal to 0, so it will be disabled / sample_every_n_epochsに0以下の値が指定されたため無効になります"
+        )
+        args.sample_every_n_epochs = None
+
+    if args.sample_every_n_steps is not None and args.sample_every_n_steps <= 0:
+        logger.warning(
+            "sample_every_n_steps is less than or equal to 0, so it will be disabled / sample_every_n_stepsに0以下の値が指定されたため無効になります"
+        )
+        args.sample_every_n_steps = None
+
 
 def add_dataset_arguments(
     parser: argparse.ArgumentParser, support_dreambooth: bool, support_caption: bool, support_caption_dropout: bool

From 691f04322a48566caf62dd67c2834ca2748c064f Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sun, 24 Mar 2024 11:10:26 +0900
Subject: [PATCH 43/69] update readme

---
 README.md | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

diff --git a/README.md b/README.md
index dd000d126..25226dff3 100644
--- a/README.md
+++ b/README.md
@@ -266,8 +266,9 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
   - Onnx may need to be updated. Onnx is not installed by default, so please install or update it with `pip install onnx==1.15.0 onnxruntime-gpu==1.17.1` etc. Please also check the comments in `requirements.txt`.
 - The model is now saved in the subdirectory as `--repo_id` in `tag_image_by_wd14_tagger.py` . This caches multiple repo_id models. Please delete unnecessary files under `--model_dir`.
 - The options `--noise_offset_random_strength` and `--ip_noise_gamma_random_strength` are added to each training script. These options can be used to vary the noise offset and ip noise gamma in the range of 0 to the specified value. PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) Thanks to KohakuBlueleaf!
-- The [English version of the dataset settings documentation](./docs/config_README-en.md) is added. PR [#1175](https://github.com/kohya-ss/sd-scripts/pull/1175) Thanks to darkstorm2150!
 - The options `--save_state_on_train_end` are added to each training script. PR [#1168](https://github.com/kohya-ss/sd-scripts/pull/1168) Thanks to gesen2egee!
+- The options `--sample_every_n_epochs` and `--sample_every_n_steps` in each training script now display a warning and ignore them when a number less than or equal to `0` is specified. Thanks to S-Del for raising the issue.
+- The [English version of the dataset settings documentation](./docs/config_README-en.md) is added. PR [#1175](https://github.com/kohya-ss/sd-scripts/pull/1175) Thanks to darkstorm2150!
 
 
 - Colab での動作時、ログ出力で停止してしまうようです。学習スクリプトに `--console_log_simple` オプションを指定し、rich のロギングを無効してお試しください。
@@ -283,8 +284,9 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
   - Onnx のバージョンアップが必要になるかもしれません。デフォルトでは Onnx はインストールされていませんので、`pip install onnx==1.15.0 onnxruntime-gpu==1.17.1` 等でインストール、アップデートしてください。`requirements.txt` のコメントもあわせてご確認ください。
 - `tag_image_by_wd14_tagger.py` で、モデルを`--repo_id` のサブディレクトリに保存するようにしました。これにより複数のモデルファイルがキャッシュされます。`--model_dir` 直下の不要なファイルは削除願います。
 - 各学習スクリプトに、noise offset、ip noise gammaを、それぞれ 0~指定した値の範囲で変動させるオプション `--noise_offset_random_strength` および `--ip_noise_gamma_random_strength` が追加されました。 PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) KohakuBlueleaf 氏に感謝します。
-- データセット設定の[英語版ドキュメント](./docs/config_README-en.md) が追加されました。PR [#1175](https://github.com/kohya-ss/sd-scripts/pull/1175) darkstorm2150 氏に感謝します。
 - 各学習スクリプトに、学習終了時に state を保存する `--save_state_on_train_end` オプションが追加されました。 PR [#1168](https://github.com/kohya-ss/sd-scripts/pull/1168) gesen2egee 氏に感謝します。
+- 各学習スクリプトで `--sample_every_n_epochs` および `--sample_every_n_steps` オプションに `0` 以下の数値を指定した時、警告を表示するとともにそれらを無視するよう変更しました。問題提起していただいた S-Del 氏に感謝します。
+- データセット設定の[英語版ドキュメント](./docs/config_README-en.md) が追加されました。PR [#1175](https://github.com/kohya-ss/sd-scripts/pull/1175) darkstorm2150 氏に感謝します。
 
 
 ### Mar 15, 2024 / 2024/3/15: v0.8.5

From 381c44955e72f04b57c74aa9b3d9a43c839c631f Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sun, 24 Mar 2024 11:27:18 +0900
Subject: [PATCH 44/69] update readme and typing hint

---
 README.md             | 2 ++
 library/train_util.py | 2 +-
 2 files changed, 3 insertions(+), 1 deletion(-)

diff --git a/README.md b/README.md
index 25226dff3..a19f7968a 100644
--- a/README.md
+++ b/README.md
@@ -254,6 +254,7 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
 
 - Colab seems to stop with log output. Try specifying `--console_log_simple` option in the training script to disable rich logging.
 - The `.toml` file for the dataset config is now read in UTF-8 encoding. PR [#1167](https://github.com/kohya-ss/sd-scripts/pull/1167) Thanks to Horizon1704!
+- Fixed a bug that the last subset settings are applied to all images when multiple subsets of regularization images are specified in the dataset settings. The settings for each subset are correctly applied to each image. PR [#1205](https://github.com/kohya-ss/sd-scripts/pull/1205) Thanks to feffy380!
 - `train_network.py` and `sdxl_train_network.py` are modified to record some dataset settings in the metadata of the trained model (`caption_prefix`, `caption_suffix`, `keep_tokens_separator`, `secondary_separator`, `enable_wildcard`).
 - Some features are added to the dataset subset settings.
   - `secondary_separator` is added to specify the tag separator that is not the target of shuffling or dropping. 
@@ -273,6 +274,7 @@ ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [docum
 
 - Colab での動作時、ログ出力で停止してしまうようです。学習スクリプトに `--console_log_simple` オプションを指定し、rich のロギングを無効してお試しください。
 - データセット設定の `.toml` ファイルが UTF-8 encoding で読み込まれるようになりました。PR [#1167](https://github.com/kohya-ss/sd-scripts/pull/1167) Horizon1704 氏に感謝します。
+- データセット設定で、正則化画像のサブセットを複数指定した時、最後のサブセットの各種設定がすべてのサブセットの画像に適用される不具合が修正されました。それぞれのサブセットの設定が、それぞれの画像に正しく適用されます。PR [#1205](https://github.com/kohya-ss/sd-scripts/pull/1205) feffy380 氏に感謝します。
 - `train_network.py` および `sdxl_train_network.py` で、学習したモデルのメタデータに一部のデータセット設定が記録されるよう修正しました（`caption_prefix`、`caption_suffix`、`keep_tokens_separator`、`secondary_separator`、`enable_wildcard`）。
 - データセットのサブセット設定にいくつかの機能を追加しました。
   - シャッフルの対象とならないタグ分割識別子の指定 `secondary_separator` を追加しました。`secondary_separator=";;;"` のように指定します。`secondary_separator` で区切ることで、その部分はシャッフル、drop 時にまとめて扱われます。
diff --git a/library/train_util.py b/library/train_util.py
index ce6e09245..99aeea90d 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -1525,7 +1525,7 @@ def load_dreambooth_dir(subset: DreamBoothSubset):
         logger.info("prepare images.")
         num_train_images = 0
         num_reg_images = 0
-        reg_infos: List[ImageInfo] = []
+        reg_infos: List[Tuple[ImageInfo, DreamBoothSubset]] = []
         for subset in subsets:
             if subset.num_repeats < 1:
                 logger.warning(

From ae97c8bfd18e4b51bdeae0a72753c8e9ceeff29d Mon Sep 17 00:00:00 2001
From: Kohaku-Blueleaf <59680068+KohakuBlueleaf@users.noreply.github.com>
Date: Sun, 24 Mar 2024 14:40:18 +0800
Subject: [PATCH 45/69] [Experimental] Add cache mechanism for dataset groups
 to avoid long waiting time for initilization (#1178)

* support meta cached dataset

* add cache meta scripts

* random ip_noise_gamma strength

* random noise_offset strength

* use correct settings for parser

* cache path/caption/size only

* revert mess up commit

* revert mess up commit

* Update requirements.txt

* Add arguments for meta cache.

* remove pickle implementation

* Return sizes when enable cache

---------

Co-authored-by: Kohya S <52813779+kohya-ss@users.noreply.github.com>
---
 cache_dataset_meta.py  | 103 +++++++++++++++++++++++++++++++++++++++++
 library/config_util.py |   4 ++
 library/train_util.py  |  83 ++++++++++++++++++++++++---------
 requirements.txt       |   2 +
 train_network.py       |   3 +-
 5 files changed, 173 insertions(+), 22 deletions(-)
 create mode 100644 cache_dataset_meta.py

diff --git a/cache_dataset_meta.py b/cache_dataset_meta.py
new file mode 100644
index 000000000..7e7d96d12
--- /dev/null
+++ b/cache_dataset_meta.py
@@ -0,0 +1,103 @@
+import argparse
+import random
+
+from accelerate.utils import set_seed
+
+import library.train_util as train_util
+import library.config_util as config_util
+from library.config_util import (
+    ConfigSanitizer,
+    BlueprintGenerator,
+)
+import library.custom_train_functions as custom_train_functions
+from library.utils import setup_logging, add_logging_arguments
+
+setup_logging()
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+def make_dataset(args):
+    train_util.prepare_dataset_args(args, True)
+    setup_logging(args, reset=True)
+
+    use_dreambooth_method = args.in_json is None
+    use_user_config = args.dataset_config is not None
+
+    if args.seed is None:
+        args.seed = random.randint(0, 2**32)
+    set_seed(args.seed)
+
+    # データセットを準備する
+    if args.dataset_class is None:
+        blueprint_generator = BlueprintGenerator(
+            ConfigSanitizer(True, True, False, True)
+        )
+        if use_user_config:
+            logger.info(f"Loading dataset config from {args.dataset_config}")
+            user_config = config_util.load_user_config(args.dataset_config)
+            ignored = ["train_data_dir", "reg_data_dir", "in_json"]
+            if any(getattr(args, attr) is not None for attr in ignored):
+                logger.warning(
+                    "ignoring the following options because config file is found: {0} / 設定ファイルが利用されるため以下のオプションは無視されます: {0}".format(
+                        ", ".join(ignored)
+                    )
+                )
+        else:
+            if use_dreambooth_method:
+                logger.info("Using DreamBooth method.")
+                user_config = {
+                    "datasets": [
+                        {
+                            "subsets": config_util.generate_dreambooth_subsets_config_by_subdirs(
+                                args.train_data_dir, args.reg_data_dir
+                            )
+                        }
+                    ]
+                }
+            else:
+                logger.info("Training with captions.")
+                user_config = {
+                    "datasets": [
+                        {
+                            "subsets": [
+                                {
+                                    "image_dir": args.train_data_dir,
+                                    "metadata_file": args.in_json,
+                                }
+                            ]
+                        }
+                    ]
+                }
+
+        blueprint = blueprint_generator.generate(user_config, args, tokenizer=None)
+        train_dataset_group = config_util.generate_dataset_group_by_blueprint(
+            blueprint.dataset_group
+        )
+    else:
+        # use arbitrary dataset class
+        train_dataset_group = train_util.load_arbitrary_dataset(args, tokenizer=None)
+    return train_dataset_group
+
+
+def setup_parser() -> argparse.ArgumentParser:
+    parser = argparse.ArgumentParser()
+    add_logging_arguments(parser)
+    train_util.add_dataset_arguments(parser, True, True, True)
+    train_util.add_training_arguments(parser, True)
+    config_util.add_config_arguments(parser)
+    custom_train_functions.add_custom_train_arguments(parser)
+    return parser
+
+
+if __name__ == "__main__":
+    parser = setup_parser()
+
+    args, unknown = parser.parse_known_args()
+    args = train_util.read_config_from_file(args, parser)
+    if args.max_token_length is None:
+        args.max_token_length = 75
+    args.cache_meta = True
+
+    dataset_group = make_dataset(args)
diff --git a/library/config_util.py b/library/config_util.py
index eb652ecf3..58ffa5f4d 100644
--- a/library/config_util.py
+++ b/library/config_util.py
@@ -111,6 +111,8 @@ class DreamBoothDatasetParams(BaseDatasetParams):
     bucket_reso_steps: int = 64
     bucket_no_upscale: bool = False
     prior_loss_weight: float = 1.0
+    cache_meta: bool = False
+    use_cached_meta: bool = False
 
 
 @dataclass
@@ -228,6 +230,8 @@ def __validate_and_convert_scalar_or_twodim(klass, value: Union[float, Sequence]
         "min_bucket_reso": int,
         "resolution": functools.partial(__validate_and_convert_scalar_or_twodim.__func__, int),
         "network_multiplier": float,
+        "cache_meta": bool,
+        "use_cached_meta": bool,
     }
 
     # options handled by argparse but not handled by user config
diff --git a/library/train_util.py b/library/train_util.py
index 99aeea90d..58c0cc14b 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -63,6 +63,7 @@
 from huggingface_hub import hf_hub_download
 import numpy as np
 from PIL import Image
+import imagesize
 import cv2
 import safetensors.torch
 from library.lpw_stable_diffusion import StableDiffusionLongPromptWeightingPipeline
@@ -1080,8 +1081,7 @@ def cache_text_encoder_outputs(
             )
 
     def get_image_size(self, image_path):
-        image = Image.open(image_path)
-        return image.size
+        return imagesize.get(image_path)
 
     def load_image_with_face_info(self, subset: BaseSubset, image_path: str):
         img = load_image(image_path)
@@ -1425,6 +1425,8 @@ def __init__(
         bucket_no_upscale: bool,
         prior_loss_weight: float,
         debug_dataset: bool,
+        cache_meta: bool,
+        use_cached_meta: bool,
     ) -> None:
         super().__init__(tokenizer, max_token_length, resolution, network_multiplier, debug_dataset)
 
@@ -1484,26 +1486,43 @@ def load_dreambooth_dir(subset: DreamBoothSubset):
                 logger.warning(f"not directory: {subset.image_dir}")
                 return [], []
 
-            img_paths = glob_images(subset.image_dir, "*")
+            sizes = None
+            if use_cached_meta:
+                logger.info(f"using cached metadata: {subset.image_dir}/dataset.txt")
+                # [img_path, caption, resolution]
+                with open(f"{subset.image_dir}/dataset.txt", "r", encoding="utf-8") as f:
+                    metas = f.readlines()
+                metas = [x.strip().split("<|##|>") for x in metas]
+                sizes = [tuple(int(res) for res in x[2].split(" ")) for x in metas]
+            
+            if use_cached_meta:
+                img_paths = [x[0] for x in metas]
+            else:
+                img_paths = glob_images(subset.image_dir, "*")
+                sizes = [None]*len(img_paths)
             logger.info(f"found directory {subset.image_dir} contains {len(img_paths)} image files")
 
-            # 画像ファイルごとにプロンプトを読み込み、もしあればそちらを使う
-            captions = []
-            missing_captions = []
-            for img_path in img_paths:
-                cap_for_img = read_caption(img_path, subset.caption_extension, subset.enable_wildcard)
-                if cap_for_img is None and subset.class_tokens is None:
-                    logger.warning(
-                        f"neither caption file nor class tokens are found. use empty caption for {img_path} / キャプションファイルもclass tokenも見つかりませんでした。空のキャプションを使用します: {img_path}"
-                    )
-                    captions.append("")
-                    missing_captions.append(img_path)
-                else:
-                    if cap_for_img is None:
-                        captions.append(subset.class_tokens)
+            if use_cached_meta:
+                captions = [x[1] for x in metas]
+                missing_captions = [x[0] for x in metas if x[1] == ""]
+            else:
+                # 画像ファイルごとにプロンプトを読み込み、もしあればそちらを使う
+                captions = []
+                missing_captions = []
+                for img_path in img_paths:
+                    cap_for_img = read_caption(img_path, subset.caption_extension, subset.enable_wildcard)
+                    if cap_for_img is None and subset.class_tokens is None:
+                        logger.warning(
+                            f"neither caption file nor class tokens are found. use empty caption for {img_path} / キャプションファイルもclass tokenも見つかりませんでした。空のキャプションを使用します: {img_path}"
+                        )
+                        captions.append("")
                         missing_captions.append(img_path)
                     else:
-                        captions.append(cap_for_img)
+                        if cap_for_img is None:
+                            captions.append(subset.class_tokens)
+                            missing_captions.append(img_path)
+                        else:
+                            captions.append(cap_for_img)
 
             self.set_tag_frequency(os.path.basename(subset.image_dir), captions)  # タグ頻度を記録
 
@@ -1520,7 +1539,21 @@ def load_dreambooth_dir(subset: DreamBoothSubset):
                         logger.warning(missing_caption + f"... and {remaining_missing_captions} more")
                         break
                     logger.warning(missing_caption)
-            return img_paths, captions
+
+            if cache_meta:
+                logger.info(f"cache metadata for {subset.image_dir}")
+                if sizes is None or sizes[0] is None:
+                    sizes = [self.get_image_size(img_path) for img_path in img_paths]
+                # [img_path, caption, resolution]
+                data = [
+                    (img_path, caption, " ".join(str(x) for x in size))
+                    for img_path, caption, size in zip(img_paths, captions, sizes)
+                ]
+                with open(f"{subset.image_dir}/dataset.txt", "w", encoding="utf-8") as f:
+                    f.write("\n".join(["<|##|>".join(x) for x in data]))
+                logger.info(f"cache metadata done for {subset.image_dir}")
+
+            return img_paths, captions, sizes
 
         logger.info("prepare images.")
         num_train_images = 0
@@ -1539,7 +1572,7 @@ def load_dreambooth_dir(subset: DreamBoothSubset):
                 )
                 continue
 
-            img_paths, captions = load_dreambooth_dir(subset)
+            img_paths, captions, sizes = load_dreambooth_dir(subset)
             if len(img_paths) < 1:
                 logger.warning(
                     f"ignore subset with image_dir='{subset.image_dir}': no images found / 画像が見つからないためサブセットを無視します"
@@ -1551,8 +1584,10 @@ def load_dreambooth_dir(subset: DreamBoothSubset):
             else:
                 num_train_images += subset.num_repeats * len(img_paths)
 
-            for img_path, caption in zip(img_paths, captions):
+            for img_path, caption, size in zip(img_paths, captions, sizes):
                 info = ImageInfo(img_path, subset.num_repeats, caption, subset.is_reg, img_path)
+                if size is not None:
+                    info.image_size = size
                 if subset.is_reg:
                     reg_infos.append((info, subset))
                 else:
@@ -3355,6 +3390,12 @@ def add_dataset_arguments(
     parser: argparse.ArgumentParser, support_dreambooth: bool, support_caption: bool, support_caption_dropout: bool
 ):
     # dataset common
+    parser.add_argument(
+        "--cache_meta", action="store_true"
+    )
+    parser.add_argument(
+        "--use_cached_meta", action="store_true"
+    )
     parser.add_argument(
         "--train_data_dir", type=str, default=None, help="directory for train images / 学習画像データのディレクトリ"
     )
diff --git a/requirements.txt b/requirements.txt
index 805f0501d..c7aeb6895 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -15,6 +15,8 @@ easygui==0.98.3
 toml==0.10.2
 voluptuous==0.13.1
 huggingface-hub==0.20.1
+# for Image utils
+imagesize==1.4.1
 # for BLIP captioning
 # requests==2.28.2
 # timm==0.6.12
diff --git a/train_network.py b/train_network.py
index 9e573d9f6..b42daba71 100644
--- a/train_network.py
+++ b/train_network.py
@@ -6,6 +6,7 @@
 import random
 import time
 import json
+import pickle
 from multiprocessing import Value
 import toml
 
@@ -23,7 +24,7 @@
 
 import library.train_util as train_util
 from library.train_util import (
-    DreamBoothDataset,
+    DreamBoothDataset, DatasetGroup
 )
 import library.config_util as config_util
 from library.config_util import (

From 025347214d761d63c5475fec83e11856f3cdbe9d Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sun, 24 Mar 2024 18:09:32 +0900
Subject: [PATCH 46/69] refactor metadata caching for DreamBooth dataset

---
 cache_dataset_meta.py    | 103 ---------------------------------------
 docs/config_README-en.md |   4 ++
 docs/config_README-ja.md |   4 ++
 library/config_util.py   |  39 +++++++++------
 library/train_util.py    |  86 ++++++++++++++++++--------------
 train_network.py         |   8 +--
 6 files changed, 85 insertions(+), 159 deletions(-)
 delete mode 100644 cache_dataset_meta.py

diff --git a/cache_dataset_meta.py b/cache_dataset_meta.py
deleted file mode 100644
index 7e7d96d12..000000000
--- a/cache_dataset_meta.py
+++ /dev/null
@@ -1,103 +0,0 @@
-import argparse
-import random
-
-from accelerate.utils import set_seed
-
-import library.train_util as train_util
-import library.config_util as config_util
-from library.config_util import (
-    ConfigSanitizer,
-    BlueprintGenerator,
-)
-import library.custom_train_functions as custom_train_functions
-from library.utils import setup_logging, add_logging_arguments
-
-setup_logging()
-import logging
-
-logger = logging.getLogger(__name__)
-
-
-def make_dataset(args):
-    train_util.prepare_dataset_args(args, True)
-    setup_logging(args, reset=True)
-
-    use_dreambooth_method = args.in_json is None
-    use_user_config = args.dataset_config is not None
-
-    if args.seed is None:
-        args.seed = random.randint(0, 2**32)
-    set_seed(args.seed)
-
-    # データセットを準備する
-    if args.dataset_class is None:
-        blueprint_generator = BlueprintGenerator(
-            ConfigSanitizer(True, True, False, True)
-        )
-        if use_user_config:
-            logger.info(f"Loading dataset config from {args.dataset_config}")
-            user_config = config_util.load_user_config(args.dataset_config)
-            ignored = ["train_data_dir", "reg_data_dir", "in_json"]
-            if any(getattr(args, attr) is not None for attr in ignored):
-                logger.warning(
-                    "ignoring the following options because config file is found: {0} / 設定ファイルが利用されるため以下のオプションは無視されます: {0}".format(
-                        ", ".join(ignored)
-                    )
-                )
-        else:
-            if use_dreambooth_method:
-                logger.info("Using DreamBooth method.")
-                user_config = {
-                    "datasets": [
-                        {
-                            "subsets": config_util.generate_dreambooth_subsets_config_by_subdirs(
-                                args.train_data_dir, args.reg_data_dir
-                            )
-                        }
-                    ]
-                }
-            else:
-                logger.info("Training with captions.")
-                user_config = {
-                    "datasets": [
-                        {
-                            "subsets": [
-                                {
-                                    "image_dir": args.train_data_dir,
-                                    "metadata_file": args.in_json,
-                                }
-                            ]
-                        }
-                    ]
-                }
-
-        blueprint = blueprint_generator.generate(user_config, args, tokenizer=None)
-        train_dataset_group = config_util.generate_dataset_group_by_blueprint(
-            blueprint.dataset_group
-        )
-    else:
-        # use arbitrary dataset class
-        train_dataset_group = train_util.load_arbitrary_dataset(args, tokenizer=None)
-    return train_dataset_group
-
-
-def setup_parser() -> argparse.ArgumentParser:
-    parser = argparse.ArgumentParser()
-    add_logging_arguments(parser)
-    train_util.add_dataset_arguments(parser, True, True, True)
-    train_util.add_training_arguments(parser, True)
-    config_util.add_config_arguments(parser)
-    custom_train_functions.add_custom_train_arguments(parser)
-    return parser
-
-
-if __name__ == "__main__":
-    parser = setup_parser()
-
-    args, unknown = parser.parse_known_args()
-    args = train_util.read_config_from_file(args, parser)
-    if args.max_token_length is None:
-        args.max_token_length = 75
-    args.cache_meta = True
-
-    dataset_group = make_dataset(args)
diff --git a/docs/config_README-en.md b/docs/config_README-en.md
index e99fde216..83bea329b 100644
--- a/docs/config_README-en.md
+++ b/docs/config_README-en.md
@@ -177,6 +177,7 @@ Options related to the configuration of DreamBooth subsets.
 | `image_dir` | `'C:\hoge'` | - | - | o (required) |
 | `caption_extension` | `".txt"` | o | o | o |
 | `class_tokens` | `"sks girl"` | - | - | o |
+| `cache_info` | `false` | o | o | o |
 | `is_reg` | `false` | - | - | o |
 
 Firstly, note that for `image_dir`, the path to the image files must be specified as being directly in the directory. Unlike the previous DreamBooth method, where images had to be placed in subdirectories, this is not compatible with that specification. Also, even if you name the folder something like "5_cat", the number of repeats of the image and the class name will not be reflected. If you want to set these individually, you will need to explicitly specify them using `num_repeats` and `class_tokens`.
@@ -187,6 +188,9 @@ Firstly, note that for `image_dir`, the path to the image files must be specifie
 * `class_tokens`
     * Sets the class tokens.
     * Only used during training when a corresponding caption file does not exist. The determination of whether or not to use it is made on a per-image basis. If `class_tokens` is not specified and a caption file is not found, an error will occur.
+* `cache_info`
+    * Specifies whether to cache the image size and caption. If not specified, it is set to `false`. The cache is saved in `metadata_cache.json` in `image_dir`.
+    * Caching speeds up the loading of the dataset after the first time. It is effective when dealing with thousands of images or more.
 * `is_reg`
     * Specifies whether the subset images are for normalization. If not specified, it is set to `false`, meaning that the images are not for normalization.
 
diff --git a/docs/config_README-ja.md b/docs/config_README-ja.md
index b57ae86a7..cc74c341b 100644
--- a/docs/config_README-ja.md
+++ b/docs/config_README-ja.md
@@ -173,6 +173,7 @@ DreamBooth 方式のサブセットの設定に関わるオプションです。
 | `image_dir` | `‘C:\hoge’` | - | - | o（必須） |
 | `caption_extension` | `".txt"` | o | o | o |
 | `class_tokens` | `“sks girl”` | - | - | o |
+| `cache_info` | `false` | o | o | o | 
 | `is_reg` | `false` | - | - | o |
 
 まず注意点として、 `image_dir` には画像ファイルが直下に置かれているパスを指定する必要があります。従来の DreamBooth の手法ではサブディレクトリに画像を置く必要がありましたが、そちらとは仕様に互換性がありません。また、`5_cat` のようなフォルダ名にしても、画像の繰り返し回数とクラス名は反映されません。これらを個別に設定したい場合、`num_repeats` と `class_tokens` で明示的に指定する必要があることに注意してください。
@@ -183,6 +184,9 @@ DreamBooth 方式のサブセットの設定に関わるオプションです。
 * `class_tokens`
     * クラストークンを設定します。
     * 画像に対応する caption ファイルが存在しない場合にのみ学習時に利用されます。利用するかどうかの判定は画像ごとに行います。`class_tokens` を指定しなかった場合に caption ファイルも見つからなかった場合にはエラーになります。
+* `cache_info`
+    * 画像サイズ、キャプションをキャッシュするかどうかを指定します。指定しなかった場合は `false` になります。キャッシュは `image_dir` に `metadata_cache.json` というファイル名で保存されます。
+    * キャッシュを行うと、二回目以降のデータセット読み込みが高速化されます。数千枚以上の画像を扱う場合には有効です。
 * `is_reg`
     * サブセットの画像が正規化用かどうかを指定します。指定しなかった場合は `false` として、つまり正規化画像ではないとして扱います。
 
diff --git a/library/config_util.py b/library/config_util.py
index 58ffa5f4d..e52b7fc02 100644
--- a/library/config_util.py
+++ b/library/config_util.py
@@ -41,12 +41,17 @@
     DatasetGroup,
 )
 from .utils import setup_logging
+
 setup_logging()
 import logging
+
 logger = logging.getLogger(__name__)
 
+
 def add_config_arguments(parser: argparse.ArgumentParser):
-    parser.add_argument("--dataset_config", type=Path, default=None, help="config file for detail settings / 詳細な設定用の設定ファイル")
+    parser.add_argument(
+        "--dataset_config", type=Path, default=None, help="config file for detail settings / 詳細な設定用の設定ファイル"
+    )
 
 
 # TODO: inherit Params class in Subset, Dataset
@@ -80,6 +85,7 @@ class DreamBoothSubsetParams(BaseSubsetParams):
     is_reg: bool = False
     class_tokens: Optional[str] = None
     caption_extension: str = ".caption"
+    cache_info: bool = False
 
 
 @dataclass
@@ -91,6 +97,7 @@ class FineTuningSubsetParams(BaseSubsetParams):
 class ControlNetSubsetParams(BaseSubsetParams):
     conditioning_data_dir: str = None
     caption_extension: str = ".caption"
+    cache_info: bool = False
 
 
 @dataclass
@@ -111,8 +118,6 @@ class DreamBoothDatasetParams(BaseDatasetParams):
     bucket_reso_steps: int = 64
     bucket_no_upscale: bool = False
     prior_loss_weight: float = 1.0
-    cache_meta: bool = False
-    use_cached_meta: bool = False
 
 
 @dataclass
@@ -202,6 +207,7 @@ def __validate_and_convert_scalar_or_twodim(klass, value: Union[float, Sequence]
     DB_SUBSET_ASCENDABLE_SCHEMA = {
         "caption_extension": str,
         "class_tokens": str,
+        "cache_info": bool,
     }
     DB_SUBSET_DISTINCT_SCHEMA = {
         Required("image_dir"): str,
@@ -214,6 +220,7 @@ def __validate_and_convert_scalar_or_twodim(klass, value: Union[float, Sequence]
     }
     CN_SUBSET_ASCENDABLE_SCHEMA = {
         "caption_extension": str,
+        "cache_info": bool,
     }
     CN_SUBSET_DISTINCT_SCHEMA = {
         Required("image_dir"): str,
@@ -230,8 +237,6 @@ def __validate_and_convert_scalar_or_twodim(klass, value: Union[float, Sequence]
         "min_bucket_reso": int,
         "resolution": functools.partial(__validate_and_convert_scalar_or_twodim.__func__, int),
         "network_multiplier": float,
-        "cache_meta": bool,
-        "use_cached_meta": bool,
     }
 
     # options handled by argparse but not handled by user config
@@ -366,7 +371,9 @@ def sanitize_argparse_namespace(self, argparse_namespace: argparse.Namespace) ->
             return self.argparse_config_validator(argparse_namespace)
         except MultipleInvalid:
             # XXX: this should be a bug
-            logger.error("Invalid cmdline parsed arguments. This should be a bug. / コマンドラインのパース結果が正しくないようです。プログラムのバグの可能性が高いです。")
+            logger.error(
+                "Invalid cmdline parsed arguments. This should be a bug. / コマンドラインのパース結果が正しくないようです。プログラムのバグの可能性が高いです。"
+            )
             raise
 
     # NOTE: value would be overwritten by latter dict if there is already the same key
@@ -551,11 +558,11 @@ def generate_dataset_group_by_blueprint(dataset_group_blueprint: DatasetGroupBlu
                     "    ",
                 )
 
-    logger.info(f'{info}')
+    logger.info(f"{info}")
 
     # make buckets first because it determines the length of dataset
     # and set the same seed for all datasets
-    seed = random.randint(0, 2**31) # actual seed is seed + epoch_no
+    seed = random.randint(0, 2**31)  # actual seed is seed + epoch_no
     for i, dataset in enumerate(datasets):
         logger.info(f"[Dataset {i}]")
         dataset.make_buckets()
@@ -642,13 +649,17 @@ def load_user_config(file: str) -> dict:
             with open(file, "r") as f:
                 config = json.load(f)
         except Exception:
-            logger.error(f"Error on parsing JSON config file. Please check the format. / JSON 形式の設定ファイルの読み込みに失敗しました。文法が正しいか確認してください。: {file}")
+            logger.error(
+                f"Error on parsing JSON config file. Please check the format. / JSON 形式の設定ファイルの読み込みに失敗しました。文法が正しいか確認してください。: {file}"
+            )
             raise
     elif file.name.lower().endswith(".toml"):
         try:
             config = toml.load(file)
         except Exception:
-            logger.error(f"Error on parsing TOML config file. Please check the format. / TOML 形式の設定ファイルの読み込みに失敗しました。文法が正しいか確認してください。: {file}")
+            logger.error(
+                f"Error on parsing TOML config file. Please check the format. / TOML 形式の設定ファイルの読み込みに失敗しました。文法が正しいか確認してください。: {file}"
+            )
             raise
     else:
         raise ValueError(f"not supported config file format / 対応していない設定ファイルの形式です: {file}")
@@ -675,13 +686,13 @@ def load_user_config(file: str) -> dict:
     train_util.prepare_dataset_args(argparse_namespace, config_args.support_finetuning)
 
     logger.info("[argparse_namespace]")
-    logger.info(f'{vars(argparse_namespace)}')
+    logger.info(f"{vars(argparse_namespace)}")
 
     user_config = load_user_config(config_args.dataset_config)
 
     logger.info("")
     logger.info("[user_config]")
-    logger.info(f'{user_config}')
+    logger.info(f"{user_config}")
 
     sanitizer = ConfigSanitizer(
         config_args.support_dreambooth, config_args.support_finetuning, config_args.support_controlnet, config_args.support_dropout
@@ -690,10 +701,10 @@ def load_user_config(file: str) -> dict:
 
     logger.info("")
     logger.info("[sanitized_user_config]")
-    logger.info(f'{sanitized_user_config}')
+    logger.info(f"{sanitized_user_config}")
 
     blueprint = BlueprintGenerator(sanitizer).generate(user_config, argparse_namespace)
 
     logger.info("")
     logger.info("[blueprint]")
-    logger.info(f'{blueprint}')
+    logger.info(f"{blueprint}")
diff --git a/library/train_util.py b/library/train_util.py
index 58c0cc14b..743a1147b 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -410,6 +410,7 @@ def __init__(
         is_reg: bool,
         class_tokens: Optional[str],
         caption_extension: str,
+        cache_info: bool,
         num_repeats,
         shuffle_caption,
         caption_separator: str,
@@ -458,6 +459,7 @@ def __init__(
         self.caption_extension = caption_extension
         if self.caption_extension and not self.caption_extension.startswith("."):
             self.caption_extension = "." + self.caption_extension
+        self.cache_info = cache_info
 
     def __eq__(self, other) -> bool:
         if not isinstance(other, DreamBoothSubset):
@@ -527,6 +529,7 @@ def __init__(
         image_dir: str,
         conditioning_data_dir: str,
         caption_extension: str,
+        cache_info: bool,
         num_repeats,
         shuffle_caption,
         caption_separator,
@@ -574,6 +577,7 @@ def __init__(
         self.caption_extension = caption_extension
         if self.caption_extension and not self.caption_extension.startswith("."):
             self.caption_extension = "." + self.caption_extension
+        self.cache_info = cache_info
 
     def __eq__(self, other) -> bool:
         if not isinstance(other, ControlNetSubset):
@@ -1410,6 +1414,8 @@ def get_item_for_caching(self, bucket, bucket_batch_size, image_index):
 
 
 class DreamBoothDataset(BaseDataset):
+    IMAGE_INFO_CACHE_FILE = "metadata_cache.json"
+
     def __init__(
         self,
         subsets: Sequence[DreamBoothSubset],
@@ -1425,8 +1431,6 @@ def __init__(
         bucket_no_upscale: bool,
         prior_loss_weight: float,
         debug_dataset: bool,
-        cache_meta: bool,
-        use_cached_meta: bool,
     ) -> None:
         super().__init__(tokenizer, max_token_length, resolution, network_multiplier, debug_dataset)
 
@@ -1486,25 +1490,36 @@ def load_dreambooth_dir(subset: DreamBoothSubset):
                 logger.warning(f"not directory: {subset.image_dir}")
                 return [], []
 
-            sizes = None
-            if use_cached_meta:
-                logger.info(f"using cached metadata: {subset.image_dir}/dataset.txt")
-                # [img_path, caption, resolution]
-                with open(f"{subset.image_dir}/dataset.txt", "r", encoding="utf-8") as f:
-                    metas = f.readlines()
-                metas = [x.strip().split("<|##|>") for x in metas]
-                sizes = [tuple(int(res) for res in x[2].split(" ")) for x in metas]
-            
-            if use_cached_meta:
-                img_paths = [x[0] for x in metas]
+            info_cache_file = os.path.join(subset.image_dir, self.IMAGE_INFO_CACHE_FILE)
+            use_cached_info_for_subset = subset.cache_info
+            if use_cached_info_for_subset:
+                logger.info(
+                    f"using cached image info for this subset / このサブセットで、キャッシュされた画像情報を使います: {info_cache_file}"
+                )
+                if not os.path.isfile(info_cache_file):
+                    logger.warning(
+                        f"image info file not found. You can ignore this warning if this is the first time to use this subset"
+                        + " / キャッシュファイルが見つかりませんでした。初回実行時はこの警告を無視してください: {metadata_file}"
+                    )
+                    use_cached_info_for_subset = False
+
+            if use_cached_info_for_subset:
+                # json: {`img_path`:{"caption": "caption...", "resolution": [width, height]}, ...}
+                with open(info_cache_file, "r", encoding="utf-8") as f:
+                    metas = json.load(f)
+                img_paths = list(metas.keys())
+                sizes = [meta["resolution"] for meta in metas.values()]
+
+                # we may need to check image size and existence of image files, but it takes time, so user should check it before training
             else:
                 img_paths = glob_images(subset.image_dir, "*")
-                sizes = [None]*len(img_paths)
+                sizes = [None] * len(img_paths)
+
             logger.info(f"found directory {subset.image_dir} contains {len(img_paths)} image files")
 
-            if use_cached_meta:
-                captions = [x[1] for x in metas]
-                missing_captions = [x[0] for x in metas if x[1] == ""]
+            if use_cached_info_for_subset:
+                captions = [meta["caption"] for meta in metas.values()]
+                missing_captions = [img_path for img_path, caption in zip(img_paths, captions) if caption is None or caption == ""]
             else:
                 # 画像ファイルごとにプロンプトを読み込み、もしあればそちらを使う
                 captions = []
@@ -1540,19 +1555,17 @@ def load_dreambooth_dir(subset: DreamBoothSubset):
                         break
                     logger.warning(missing_caption)
 
-            if cache_meta:
-                logger.info(f"cache metadata for {subset.image_dir}")
-                if sizes is None or sizes[0] is None:
-                    sizes = [self.get_image_size(img_path) for img_path in img_paths]
-                # [img_path, caption, resolution]
-                data = [
-                    (img_path, caption, " ".join(str(x) for x in size))
-                    for img_path, caption, size in zip(img_paths, captions, sizes)
-                ]
-                with open(f"{subset.image_dir}/dataset.txt", "w", encoding="utf-8") as f:
-                    f.write("\n".join(["<|##|>".join(x) for x in data]))
-                logger.info(f"cache metadata done for {subset.image_dir}")
-
+            if not use_cached_info_for_subset and subset.cache_info:
+                logger.info(f"cache image info for / 画像情報をキャッシュします : {info_cache_file}")
+                sizes = [self.get_image_size(img_path) for img_path in tqdm(img_paths, desc="get image size")]
+                matas = {}
+                for img_path, caption, size in zip(img_paths, captions, sizes):
+                    matas[img_path] = {"caption": caption, "resolution": list(size)}
+                with open(info_cache_file, "w", encoding="utf-8") as f:
+                    json.dump(matas, f, ensure_ascii=False, indent=2)
+                logger.info(f"cache image info done for / 画像情報を出力しました : {info_cache_file}")
+
+            # if sizes are not set, image size will be read in make_buckets
             return img_paths, captions, sizes
 
         logger.info("prepare images.")
@@ -1873,7 +1886,8 @@ def __init__(
                 subset.image_dir,
                 False,
                 None,
-                subset.caption_extension,
+                subset.caption_extension, 
+                subset.cache_info,
                 subset.num_repeats,
                 subset.shuffle_caption,
                 subset.caption_separator,
@@ -3391,13 +3405,13 @@ def add_dataset_arguments(
 ):
     # dataset common
     parser.add_argument(
-        "--cache_meta", action="store_true"
-    )
-    parser.add_argument(
-        "--use_cached_meta", action="store_true"
+        "--train_data_dir", type=str, default=None, help="directory for train images / 学習画像データのディレクトリ"
     )
     parser.add_argument(
-        "--train_data_dir", type=str, default=None, help="directory for train images / 学習画像データのディレクトリ"
+        "--cache_info",
+        action="store_true",
+        help="cache meta information (caption and image size) for faster dataset loading. only available for DreamBooth"
+        + " / メタ情報（キャプションとサイズ）をキャッシュしてデータセット読み込みを高速化する。DreamBooth方式のみ有効",
     )
     parser.add_argument(
         "--shuffle_caption", action="store_true", help="shuffle separated caption / 区切られたcaptionの各要素をshuffleする"
diff --git a/train_network.py b/train_network.py
index b42daba71..7ae9283cb 100644
--- a/train_network.py
+++ b/train_network.py
@@ -6,7 +6,6 @@
 import random
 import time
 import json
-import pickle
 from multiprocessing import Value
 import toml
 
@@ -14,18 +13,15 @@
 
 import torch
 from library.device_utils import init_ipex, clean_memory_on_device
-init_ipex()
 
-from torch.nn.parallel import DistributedDataParallel as DDP
+init_ipex()
 
 from accelerate.utils import set_seed
 from diffusers import DDPMScheduler
 from library import model_util
 
 import library.train_util as train_util
-from library.train_util import (
-    DreamBoothDataset, DatasetGroup
-)
+from library.train_util import DreamBoothDataset
 import library.config_util as config_util
 from library.config_util import (
     ConfigSanitizer,

From 1648ade6da549c7def2e21f236453e7938c499cd Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sun, 24 Mar 2024 20:55:48 +0900
Subject: [PATCH 47/69] format by black

---
 library/config_util.py | 31 +++++++++++++++++++++----------
 1 file changed, 21 insertions(+), 10 deletions(-)

diff --git a/library/config_util.py b/library/config_util.py
index eb652ecf3..ff4de0921 100644
--- a/library/config_util.py
+++ b/library/config_util.py
@@ -41,12 +41,17 @@
     DatasetGroup,
 )
 from .utils import setup_logging
+
 setup_logging()
 import logging
+
 logger = logging.getLogger(__name__)
 
+
 def add_config_arguments(parser: argparse.ArgumentParser):
-    parser.add_argument("--dataset_config", type=Path, default=None, help="config file for detail settings / 詳細な設定用の設定ファイル")
+    parser.add_argument(
+        "--dataset_config", type=Path, default=None, help="config file for detail settings / 詳細な設定用の設定ファイル"
+    )
 
 
 # TODO: inherit Params class in Subset, Dataset
@@ -362,7 +367,9 @@ def sanitize_argparse_namespace(self, argparse_namespace: argparse.Namespace) ->
             return self.argparse_config_validator(argparse_namespace)
         except MultipleInvalid:
             # XXX: this should be a bug
-            logger.error("Invalid cmdline parsed arguments. This should be a bug. / コマンドラインのパース結果が正しくないようです。プログラムのバグの可能性が高いです。")
+            logger.error(
+                "Invalid cmdline parsed arguments. This should be a bug. / コマンドラインのパース結果が正しくないようです。プログラムのバグの可能性が高いです。"
+            )
             raise
 
     # NOTE: value would be overwritten by latter dict if there is already the same key
@@ -547,11 +554,11 @@ def generate_dataset_group_by_blueprint(dataset_group_blueprint: DatasetGroupBlu
                     "    ",
                 )
 
-    logger.info(f'{info}')
+    logger.info(f"{info}")
 
     # make buckets first because it determines the length of dataset
     # and set the same seed for all datasets
-    seed = random.randint(0, 2**31) # actual seed is seed + epoch_no
+    seed = random.randint(0, 2**31)  # actual seed is seed + epoch_no
     for i, dataset in enumerate(datasets):
         logger.info(f"[Dataset {i}]")
         dataset.make_buckets()
@@ -638,13 +645,17 @@ def load_user_config(file: str) -> dict:
             with open(file, "r") as f:
                 config = json.load(f)
         except Exception:
-            logger.error(f"Error on parsing JSON config file. Please check the format. / JSON 形式の設定ファイルの読み込みに失敗しました。文法が正しいか確認してください。: {file}")
+            logger.error(
+                f"Error on parsing JSON config file. Please check the format. / JSON 形式の設定ファイルの読み込みに失敗しました。文法が正しいか確認してください。: {file}"
+            )
             raise
     elif file.name.lower().endswith(".toml"):
         try:
             config = toml.load(file)
         except Exception:
-            logger.error(f"Error on parsing TOML config file. Please check the format. / TOML 形式の設定ファイルの読み込みに失敗しました。文法が正しいか確認してください。: {file}")
+            logger.error(
+                f"Error on parsing TOML config file. Please check the format. / TOML 形式の設定ファイルの読み込みに失敗しました。文法が正しいか確認してください。: {file}"
+            )
             raise
     else:
         raise ValueError(f"not supported config file format / 対応していない設定ファイルの形式です: {file}")
@@ -671,13 +682,13 @@ def load_user_config(file: str) -> dict:
     train_util.prepare_dataset_args(argparse_namespace, config_args.support_finetuning)
 
     logger.info("[argparse_namespace]")
-    logger.info(f'{vars(argparse_namespace)}')
+    logger.info(f"{vars(argparse_namespace)}")
 
     user_config = load_user_config(config_args.dataset_config)
 
     logger.info("")
     logger.info("[user_config]")
-    logger.info(f'{user_config}')
+    logger.info(f"{user_config}")
 
     sanitizer = ConfigSanitizer(
         config_args.support_dreambooth, config_args.support_finetuning, config_args.support_controlnet, config_args.support_dropout
@@ -686,10 +697,10 @@ def load_user_config(file: str) -> dict:
 
     logger.info("")
     logger.info("[sanitized_user_config]")
-    logger.info(f'{sanitized_user_config}')
+    logger.info(f"{sanitized_user_config}")
 
     blueprint = BlueprintGenerator(sanitizer).generate(user_config, argparse_namespace)
 
     logger.info("")
     logger.info("[blueprint]")
-    logger.info(f'{blueprint}')
+    logger.info(f"{blueprint}")

From 9bbb28c3619a9ff86a51bdc7ea83584976840663 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sun, 24 Mar 2024 22:06:37 +0900
Subject: [PATCH 48/69] update PyTorch version and reorganize dependencies

---
 README-ja.md                    |  55 ++----------
 README.md                       | 149 ++++----------------------------
 docs/train_SDXL-en.md           |  84 ++++++++++++++++++
 requirements.txt                |   8 +-
 sdxl_minimal_inference.py       |  30 +++++--
 sdxl_train_textual_inversion.py |   1 -
 6 files changed, 136 insertions(+), 191 deletions(-)
 create mode 100644 docs/train_SDXL-en.md

diff --git a/README-ja.md b/README-ja.md
index 29c33a659..1d83c44f1 100644
--- a/README-ja.md
+++ b/README-ja.md
@@ -1,7 +1,3 @@
-SDXLがサポートされました。sdxlブランチはmainブランチにマージされました。リポジトリを更新したときにはUpgradeの手順を実行してください。また accelerate のバージョンが上がっていますので、accelerate config を再度実行してください。
-
-SDXL学習については[こちら](./README.md#sdxl-training)をご覧ください（英語です）。
-
 ## リポジトリについて
 Stable Diffusionの学習、画像生成、その他のスクリプトを入れたリポジトリです。
 
@@ -21,6 +17,7 @@ GUIやPowerShellスクリプトなど、より使いやすくする機能が[bma
 
 * [学習について、共通編](./docs/train_README-ja.md) : データ整備やオプションなど
     * [データセット設定](./docs/config_README-ja.md)
+* [SDXL学習](./docs/train_SDXL-en.md) （英語版）
 * [DreamBoothの学習について](./docs/train_db_README-ja.md)
 * [fine-tuningのガイド](./docs/fine_tune_README_ja.md):
 * [LoRAの学習について](./docs/train_network_README-ja.md)
@@ -44,9 +41,7 @@ PowerShellを使う場合、venvを使えるようにするためには以下の
 
 ## Windows環境でのインストール
 
-スクリプトはPyTorch 2.0.1でテストしています。PyTorch 1.12.1でも動作すると思われます。
-
-以下の例ではPyTorchは2.0.1／CUDA 11.8版をインストールします。CUDA 11.6版やPyTorch 1.12.1を使う場合は適宜書き換えください。
+スクリプトはPyTorch 2.1.1でテストしています。PyTorch 2.0.1、1.12.1でも動作すると思われます。
 
 （なお、python -m venv～の行で「python」とだけ表示された場合、py -m venv～のようにpythonをpyに変更してください。）
 
@@ -59,20 +54,20 @@ cd sd-scripts
 python -m venv venv
 .\venv\Scripts\activate
 
-pip install torch==2.0.1+cu118 torchvision==0.15.2+cu118 --index-url https://download.pytorch.org/whl/cu118
+pip install torch==2.1.1 torchvision==0.16.1 --index-url https://download.pytorch.org/whl/cu118
 pip install --upgrade -r requirements.txt
-pip install xformers==0.0.20
+pip install xformers==0.0.23 --index-url https://download.pytorch.org/whl/cu118
 
 accelerate config
 ```
 
 コマンドプロンプトでも同一です。
 
-（注:``python -m venv venv`` のほうが ``python -m venv --system-site-packages venv`` より安全そうなため書き換えました。globalなpythonにパッケージがインストールしてあると、後者だといろいろと問題が起きます。）
+注：`bitsandbytes==0.43.0`、`prodigyopt==1.0`、`lion-pytorch==0.0.6` は `requirements.txt` に含まれるようになりました。他のバージョンを使う場合は適宜インストールしてください。
 
-accelerate configの質問には以下のように答えてください。（bf16で学習する場合、最後の質問にはbf16と答えてください。）
+この例では PyTorch および xfomers は2.1.1／CUDA 11.8版をインストールします。CUDA 12.1版やPyTorch 1.12.1を使う場合は適宜書き換えください。たとえば CUDA 12.1版の場合は `pip install torch==2.1.1 torchvision==0.16.1 --index-url https://download.pytorch.org/whl/cu121` および `pip install xformers==0.0.23 --index-url https://download.pytorch.org/whl/cu121` としてください。
 
-※0.15.0から日本語環境では選択のためにカーソルキーを押すと落ちます（……）。数字キーの0、1、2……で選択できますので、そちらを使ってください。
+accelerate configの質問には以下のように答えてください。（bf16で学習する場合、最後の質問にはbf16と答えてください。）
 
 ```txt
 - This machine
@@ -87,41 +82,6 @@ accelerate configの質問には以下のように答えてください。（bf1
 ※場合によって ``ValueError: fp16 mixed precision requires a GPU`` というエラーが出ることがあるようです。この場合、6番目の質問（
 ``What GPU(s) (by id) should be used for training on this machine as a comma-separated list? [all]:``）に「0」と答えてください。（id `0`のGPUが使われます。）
 
-### オプション：`bitsandbytes`（8bit optimizer）を使う
-
-`bitsandbytes`はオプションになりました。Linuxでは通常通りpipでインストールできます（0.41.1または以降のバージョンを推奨）。
-
-Windowsでは0.35.0または0.41.1を推奨します。
-
-- `bitsandbytes` 0.35.0: 安定しているとみられるバージョンです。AdamW8bitは使用できますが、他のいくつかの8bit optimizer、学習時の`full_bf16`オプションは使用できません。
-- `bitsandbytes` 0.41.1: Lion8bit、PagedAdamW8bit、PagedLion8bitをサポートします。`full_bf16`が使用できます。
-
-注：`bitsandbytes` 0.35.0から0.41.0までのバージョンには問題があるようです。 https://github.com/TimDettmers/bitsandbytes/issues/659
-
-以下の手順に従い、`bitsandbytes`をインストールしてください。
-
-### 0.35.0を使う場合
-
-PowerShellの例です。コマンドプロンプトではcpの代わりにcopyを使ってください。
-
-```powershell
-cd sd-scripts
-.\venv\Scripts\activate
-pip install bitsandbytes==0.35.0
-
-cp .\bitsandbytes_windows\*.dll .\venv\Lib\site-packages\bitsandbytes\
-cp .\bitsandbytes_windows\cextension.py .\venv\Lib\site-packages\bitsandbytes\cextension.py
-cp .\bitsandbytes_windows\main.py .\venv\Lib\site-packages\bitsandbytes\cuda_setup\main.py
-```
-
-### 0.41.1を使う場合
-
-jllllll氏の配布されている[こちら](https://github.com/jllllll/bitsandbytes-windows-webui) または他の場所から、Windows用のwhlファイルをインストールしてください。
-
-```powershell
-python -m pip install bitsandbytes==0.41.1 --prefer-binary --extra-index-url=https://jllllll.github.io/bitsandbytes-windows-webui
-```
-
 ## アップグレード
 
 新しいリリースがあった場合、以下のコマンドで更新できます。
@@ -151,4 +111,3 @@ Conv2d 3x3への拡大は [cloneofsimo氏](https://github.com/cloneofsimo/lora)
 
 [BLIP](https://github.com/salesforce/BLIP): BSD-3-Clause
 
-
diff --git a/README.md b/README.md
index a19f7968a..ef26acab8 100644
--- a/README.md
+++ b/README.md
@@ -1,5 +1,3 @@
-__SDXL is now supported. The sdxl branch has been merged into the main branch. If you update the repository, please follow the upgrade instructions. Also, the version of accelerate has been updated, so please run accelerate config again.__ The documentation for SDXL training is [here](./README.md#sdxl-training).
-
 This repository contains training, generation and utility scripts for Stable Diffusion.
 
 [__Change History__](#change-history) is moved to the bottom of the page. 
@@ -20,9 +18,9 @@ This repository contains the scripts for:
 
 ## About requirements.txt
 
-These files do not contain requirements for PyTorch. Because the versions of them depend on your environment. Please install PyTorch at first (see installation guide below.) 
+The file does not contain requirements for PyTorch. Because the version of PyTorch depends on the environment, it is not included in the file. Please install PyTorch first according to the environment. See installation instructions below.
 
-The scripts are tested with Pytorch 2.0.1. 1.12.1 is not tested but should work.
+The scripts are tested with Pytorch 2.1.1. 2.0.1 and 1.12.1 is not tested but should work.
 
 ## Links to usage documentation
 
@@ -32,12 +30,13 @@ Most of the documents are written in Japanese.
 
 * [Training guide - common](./docs/train_README-ja.md) : data preparation, options etc... 
   * [Chinese version](./docs/train_README-zh.md)
+* [SDXL training](./docs/train_SDXL-en.md) (English version)
 * [Dataset config](./docs/config_README-ja.md) 
   * [English version](./docs/config_README-en.md)
 * [DreamBooth training guide](./docs/train_db_README-ja.md)
 * [Step by Step fine-tuning guide](./docs/fine_tune_README_ja.md):
-* [training LoRA](./docs/train_network_README-ja.md)
-* [training Textual Inversion](./docs/train_ti_README-ja.md)
+* [Training LoRA](./docs/train_network_README-ja.md)
+* [Training Textual Inversion](./docs/train_ti_README-ja.md)
 * [Image generation](./docs/gen_img_README-ja.md)
 * note.com [Model conversion](https://note.com/kohya_ss/n/n374f316fe4ad)
 
@@ -65,14 +64,18 @@ cd sd-scripts
 python -m venv venv
 .\venv\Scripts\activate
 
-pip install torch==2.0.1+cu118 torchvision==0.15.2+cu118 --index-url https://download.pytorch.org/whl/cu118
+pip install torch==2.1.1 torchvision==0.16.1 --index-url https://download.pytorch.org/whl/cu118
 pip install --upgrade -r requirements.txt
-pip install xformers==0.0.20
+pip install xformers==0.0.23 --index-url https://download.pytorch.org/whl/cu118
 
 accelerate config
 ```
 
-__Note:__ Now bitsandbytes is optional. Please install any version of bitsandbytes as needed. Installation instructions are in the following section.
+If `python -m venv` shows only `python`, change `python` to `py`.
+
+__Note:__ Now `bitsandbytes==0.43.0`, `prodigyopt==1.0` and `lion-pytorch==0.0.6` are included in the requirements.txt. If you'd like to use the another version, please install it manually.
+
+This installation is for CUDA 11.8. If you use a different version of CUDA, please install the appropriate version of PyTorch and xformers. For example, if you use CUDA 12, please install `pip install torch==2.1.1 torchvision==0.16.1 --index-url https://download.pytorch.org/whl/cu121` and `pip install xformers==0.0.23 --index-url https://download.pytorch.org/whl/cu121`.
 
 <!-- 
 cp .\bitsandbytes_windows\*.dll .\venv\Lib\site-packages\bitsandbytes\
@@ -91,48 +94,13 @@ Answers to accelerate config:
 - fp16
 ```
 
-note: Some user reports ``ValueError: fp16 mixed precision requires a GPU`` is occurred in training. In this case, answer `0` for the 6th question: 
+If you'd like to use bf16, please answer `bf16` to the last question.
+
+Note: Some user reports ``ValueError: fp16 mixed precision requires a GPU`` is occurred in training. In this case, answer `0` for the 6th question: 
 ``What GPU(s) (by id) should be used for training on this machine as a comma-separated list? [all]:`` 
 
 (Single GPU with id `0` will be used.)
 
-### Optional: Use `bitsandbytes` (8bit optimizer)
-
-For 8bit optimizer, you need to install `bitsandbytes`. For Linux, please install `bitsandbytes` as usual (0.41.1 or later is recommended.)
-
-For Windows, there are several versions of `bitsandbytes`:
-
-- `bitsandbytes` 0.35.0: Stable version. AdamW8bit is available. `full_bf16` is not available.
-- `bitsandbytes` 0.41.1: Lion8bit, PagedAdamW8bit and PagedLion8bit are available. `full_bf16` is available.
-
-Note: `bitsandbytes`above 0.35.0 till 0.41.0 seems to have an issue: https://github.com/TimDettmers/bitsandbytes/issues/659
-
-Follow the instructions below to install `bitsandbytes` for Windows.
-
-### bitsandbytes 0.35.0 for Windows
-
-Open a regular Powershell terminal and type the following inside:
-
-```powershell
-cd sd-scripts
-.\venv\Scripts\activate
-pip install bitsandbytes==0.35.0
-
-cp .\bitsandbytes_windows\*.dll .\venv\Lib\site-packages\bitsandbytes\
-cp .\bitsandbytes_windows\cextension.py .\venv\Lib\site-packages\bitsandbytes\cextension.py
-cp .\bitsandbytes_windows\main.py .\venv\Lib\site-packages\bitsandbytes\cuda_setup\main.py
-```
-
-This will install `bitsandbytes` 0.35.0 and copy the necessary files to the `bitsandbytes` directory.
-
-### bitsandbytes 0.41.1 for Windows
-
-Install the Windows version whl file from [here](https://github.com/jllllll/bitsandbytes-windows-webui) or other sources, like:
-
-```powershell
-python -m pip install bitsandbytes==0.41.1 --prefer-binary --extra-index-url=https://jllllll.github.io/bitsandbytes-windows-webui
-```
-
 ## Upgrade
 
 When a new release comes out you can upgrade your repo with the following command:
@@ -163,91 +131,6 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
 [BLIP](https://github.com/salesforce/BLIP): BSD-3-Clause
 
 
-## SDXL training
-
-The documentation in this section will be moved to a separate document later.
-
-### Training scripts for SDXL
-
-- `sdxl_train.py` is a script for SDXL fine-tuning. The usage is almost the same as `fine_tune.py`, but it also supports DreamBooth dataset.
-  - `--full_bf16` option is added. Thanks to KohakuBlueleaf!
-    - This option enables the full bfloat16 training (includes gradients). This option is useful to reduce the GPU memory usage. 
-    - The full bfloat16 training might be unstable. Please use it at your own risk.
-  - The different learning rates for each U-Net block are now supported in sdxl_train.py. Specify with `--block_lr` option. Specify 23 values separated by commas like `--block_lr 1e-3,1e-3 ... 1e-3`.
-    - 23 values correspond to `0: time/label embed, 1-9: input blocks 0-8, 10-12: mid blocks 0-2, 13-21: output blocks 0-8, 22: out`.
-- `prepare_buckets_latents.py` now supports SDXL fine-tuning.
-
-- `sdxl_train_network.py` is a script for LoRA training for SDXL. The usage is almost the same as `train_network.py`.
-
-- Both scripts has following additional options:
-  - `--cache_text_encoder_outputs` and `--cache_text_encoder_outputs_to_disk`: Cache the outputs of the text encoders. This option is useful to reduce the GPU memory usage. This option cannot be used with options for shuffling or dropping the captions.
-  - `--no_half_vae`: Disable the half-precision (mixed-precision) VAE. VAE for SDXL seems to produce NaNs in some cases. This option is useful to avoid the NaNs.
-
-- `--weighted_captions` option is not supported yet for both scripts.
-
-- `sdxl_train_textual_inversion.py` is a script for Textual Inversion training for SDXL. The usage is almost the same as `train_textual_inversion.py`.
-  - `--cache_text_encoder_outputs` is not supported.
-  - There are two options for captions:
-    1. Training with captions. All captions must include the token string. The token string is replaced with multiple tokens.
-    2. Use `--use_object_template` or `--use_style_template` option. The captions are generated from the template. The existing captions are ignored.
-  - See below for the format of the embeddings.
-
-- `--min_timestep` and `--max_timestep` options are added to each training script. These options can be used to train U-Net with different timesteps. The default values are 0 and 1000.
-
-### Utility scripts for SDXL
-
-- `tools/cache_latents.py` is added. This script can be used to cache the latents to disk in advance. 
-  - The options are almost the same as `sdxl_train.py'. See the help message for the usage.
-  - Please launch the script as follows:
-    `accelerate launch  --num_cpu_threads_per_process 1 tools/cache_latents.py ...`
-  - This script should work with multi-GPU, but it is not tested in my environment.
-
-- `tools/cache_text_encoder_outputs.py` is added. This script can be used to cache the text encoder outputs to disk in advance. 
-  - The options are almost the same as `cache_latents.py` and `sdxl_train.py`. See the help message for the usage.
-
-- `sdxl_gen_img.py` is added. This script can be used to generate images with SDXL, including LoRA, Textual Inversion and ControlNet-LLLite. See the help message for the usage.
-
-### Tips for SDXL training
-
-- The default resolution of SDXL is 1024x1024.
-- The fine-tuning can be done with 24GB GPU memory with the batch size of 1. For 24GB GPU, the following options are recommended __for the fine-tuning with 24GB GPU memory__:
-  - Train U-Net only.
-  - Use gradient checkpointing.
-  - Use `--cache_text_encoder_outputs` option and caching latents.
-  - Use Adafactor optimizer. RMSprop 8bit or Adagrad 8bit may work. AdamW 8bit doesn't seem to work.
-- The LoRA training can be done with 8GB GPU memory (10GB recommended). For reducing the GPU memory usage, the following options are recommended:
-  - Train U-Net only.
-  - Use gradient checkpointing.
-  - Use `--cache_text_encoder_outputs` option and caching latents.
-  - Use one of 8bit optimizers or Adafactor optimizer.
-  - Use lower dim (4 to 8 for 8GB GPU).
-- `--network_train_unet_only` option is highly recommended for SDXL LoRA. Because SDXL has two text encoders, the result of the training will be unexpected.
-- PyTorch 2 seems to use slightly less GPU memory than PyTorch 1.
-- `--bucket_reso_steps` can be set to 32 instead of the default value 64. Smaller values than 32 will not work for SDXL training.
-
-Example of the optimizer settings for Adafactor with the fixed learning rate:
-```toml
-optimizer_type = "adafactor"
-optimizer_args = [ "scale_parameter=False", "relative_step=False", "warmup_init=False" ]
-lr_scheduler = "constant_with_warmup"
-lr_warmup_steps = 100
-learning_rate = 4e-7 # SDXL original learning rate
-```
-
-### Format of Textual Inversion embeddings for SDXL
-
-```python
-from safetensors.torch import save_file
-
-state_dict = {"clip_g": embs_for_text_encoder_1280, "clip_l": embs_for_text_encoder_768}
-save_file(state_dict, file)
-```
-
-### ControlNet-LLLite
-
-ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [documentation](./docs/train_lllite_README.md) for details.
-
-
 ## Change History
 
 ### Working in progress
@@ -362,6 +245,8 @@ We would like to express our deep gratitude to Mark Saint (cacoe) from leonardo.
 Please read [Releases](https://github.com/kohya-ss/sd-scripts/releases) for recent updates.
 最近の更新情報は [Release](https://github.com/kohya-ss/sd-scripts/releases) をご覧ください。
 
+## Additional Information
+
 ### Naming of LoRA
 
 The LoRA supported by `train_network.py` has been named to avoid confusion. The documentation has been updated. The following are the names of LoRA types in this repository.
diff --git a/docs/train_SDXL-en.md b/docs/train_SDXL-en.md
new file mode 100644
index 000000000..a4c55b3fd
--- /dev/null
+++ b/docs/train_SDXL-en.md
@@ -0,0 +1,84 @@
+## SDXL training
+
+The documentation will be moved to the training documentation in the future. The following is a brief explanation of the training scripts for SDXL.
+
+### Training scripts for SDXL
+
+- `sdxl_train.py` is a script for SDXL fine-tuning. The usage is almost the same as `fine_tune.py`, but it also supports DreamBooth dataset.
+  - `--full_bf16` option is added. Thanks to KohakuBlueleaf!
+    - This option enables the full bfloat16 training (includes gradients). This option is useful to reduce the GPU memory usage. 
+    - The full bfloat16 training might be unstable. Please use it at your own risk.
+  - The different learning rates for each U-Net block are now supported in sdxl_train.py. Specify with `--block_lr` option. Specify 23 values separated by commas like `--block_lr 1e-3,1e-3 ... 1e-3`.
+    - 23 values correspond to `0: time/label embed, 1-9: input blocks 0-8, 10-12: mid blocks 0-2, 13-21: output blocks 0-8, 22: out`.
+- `prepare_buckets_latents.py` now supports SDXL fine-tuning.
+
+- `sdxl_train_network.py` is a script for LoRA training for SDXL. The usage is almost the same as `train_network.py`.
+
+- Both scripts has following additional options:
+  - `--cache_text_encoder_outputs` and `--cache_text_encoder_outputs_to_disk`: Cache the outputs of the text encoders. This option is useful to reduce the GPU memory usage. This option cannot be used with options for shuffling or dropping the captions.
+  - `--no_half_vae`: Disable the half-precision (mixed-precision) VAE. VAE for SDXL seems to produce NaNs in some cases. This option is useful to avoid the NaNs.
+
+- `--weighted_captions` option is not supported yet for both scripts.
+
+- `sdxl_train_textual_inversion.py` is a script for Textual Inversion training for SDXL. The usage is almost the same as `train_textual_inversion.py`.
+  - `--cache_text_encoder_outputs` is not supported.
+  - There are two options for captions:
+    1. Training with captions. All captions must include the token string. The token string is replaced with multiple tokens.
+    2. Use `--use_object_template` or `--use_style_template` option. The captions are generated from the template. The existing captions are ignored.
+  - See below for the format of the embeddings.
+
+- `--min_timestep` and `--max_timestep` options are added to each training script. These options can be used to train U-Net with different timesteps. The default values are 0 and 1000.
+
+### Utility scripts for SDXL
+
+- `tools/cache_latents.py` is added. This script can be used to cache the latents to disk in advance. 
+  - The options are almost the same as `sdxl_train.py'. See the help message for the usage.
+  - Please launch the script as follows:
+    `accelerate launch  --num_cpu_threads_per_process 1 tools/cache_latents.py ...`
+  - This script should work with multi-GPU, but it is not tested in my environment.
+
+- `tools/cache_text_encoder_outputs.py` is added. This script can be used to cache the text encoder outputs to disk in advance. 
+  - The options are almost the same as `cache_latents.py` and `sdxl_train.py`. See the help message for the usage.
+
+- `sdxl_gen_img.py` is added. This script can be used to generate images with SDXL, including LoRA, Textual Inversion and ControlNet-LLLite. See the help message for the usage.
+
+### Tips for SDXL training
+
+- The default resolution of SDXL is 1024x1024.
+- The fine-tuning can be done with 24GB GPU memory with the batch size of 1. For 24GB GPU, the following options are recommended __for the fine-tuning with 24GB GPU memory__:
+  - Train U-Net only.
+  - Use gradient checkpointing.
+  - Use `--cache_text_encoder_outputs` option and caching latents.
+  - Use Adafactor optimizer. RMSprop 8bit or Adagrad 8bit may work. AdamW 8bit doesn't seem to work.
+- The LoRA training can be done with 8GB GPU memory (10GB recommended). For reducing the GPU memory usage, the following options are recommended:
+  - Train U-Net only.
+  - Use gradient checkpointing.
+  - Use `--cache_text_encoder_outputs` option and caching latents.
+  - Use one of 8bit optimizers or Adafactor optimizer.
+  - Use lower dim (4 to 8 for 8GB GPU).
+- `--network_train_unet_only` option is highly recommended for SDXL LoRA. Because SDXL has two text encoders, the result of the training will be unexpected.
+- PyTorch 2 seems to use slightly less GPU memory than PyTorch 1.
+- `--bucket_reso_steps` can be set to 32 instead of the default value 64. Smaller values than 32 will not work for SDXL training.
+
+Example of the optimizer settings for Adafactor with the fixed learning rate:
+```toml
+optimizer_type = "adafactor"
+optimizer_args = [ "scale_parameter=False", "relative_step=False", "warmup_init=False" ]
+lr_scheduler = "constant_with_warmup"
+lr_warmup_steps = 100
+learning_rate = 4e-7 # SDXL original learning rate
+```
+
+### Format of Textual Inversion embeddings for SDXL
+
+```python
+from safetensors.torch import save_file
+
+state_dict = {"clip_g": embs_for_text_encoder_1280, "clip_l": embs_for_text_encoder_768}
+save_file(state_dict, file)
+```
+
+### ControlNet-LLLite
+
+ControlNet-LLLite, a novel method for ControlNet with SDXL, is added. See [documentation](./docs/train_lllite_README.md) for details.
+
diff --git a/requirements.txt b/requirements.txt
index 805f0501d..51085744e 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -6,8 +6,10 @@ ftfy==6.1.1
 opencv-python==4.7.0.68
 einops==0.7.0
 pytorch-lightning==1.9.0
-# bitsandbytes==0.39.1
-tensorboard==2.10.1
+bitsandbytes==0.43.0
+prodigyopt==1.0
+lion-pytorch==0.0.6
+tensorboard
 safetensors==0.4.2
 # gradio==3.16.2
 altair==4.2.2
@@ -31,7 +33,7 @@ huggingface-hub==0.20.1
 # this is for onnx: 
 # protobuf==3.20.3
 # open clip for SDXL
-open-clip-torch==2.20.0
+# open-clip-torch==2.20.0
 # For logging
 rich==13.7.0
 # for kohya_ss library
diff --git a/sdxl_minimal_inference.py b/sdxl_minimal_inference.py
index 084735665..a1e93b7f0 100644
--- a/sdxl_minimal_inference.py
+++ b/sdxl_minimal_inference.py
@@ -11,20 +11,24 @@
 
 import torch
 from library.device_utils import init_ipex, get_preferred_device
+
 init_ipex()
 
 from tqdm import tqdm
 from transformers import CLIPTokenizer
 from diffusers import EulerDiscreteScheduler
 from PIL import Image
-import open_clip
+
+# import open_clip
 from safetensors.torch import load_file
 
 from library import model_util, sdxl_model_util
 import networks.lora as lora
 from library.utils import setup_logging
+
 setup_logging()
 import logging
+
 logger = logging.getLogger(__name__)
 
 # scheduler: このあたりの設定はSD1/2と同じでいいらしい
@@ -154,12 +158,13 @@ def get_timestep_embedding(x, outdim):
     text_model2.eval()
 
     unet.set_use_memory_efficient_attention(True, False)
-    if torch.__version__ >= "2.0.0": # PyTorch 2.0.0 以上対応のxformersなら以下が使える
+    if torch.__version__ >= "2.0.0":  # PyTorch 2.0.0 以上対応のxformersなら以下が使える
         vae.set_use_memory_efficient_attention_xformers(True)
 
     # Tokenizers
     tokenizer1 = CLIPTokenizer.from_pretrained(text_encoder_1_name)
-    tokenizer2 = lambda x: open_clip.tokenize(x, context_length=77)
+    # tokenizer2 = lambda x: open_clip.tokenize(x, context_length=77)
+    tokenizer2 = CLIPTokenizer.from_pretrained(text_encoder_2_name)
 
     # LoRA
     for weights_file in args.lora_weights:
@@ -192,7 +197,9 @@ def generate_image(prompt, prompt2, negative_prompt, seed=None):
             emb3 = get_timestep_embedding(torch.FloatTensor([target_height, target_width]).unsqueeze(0), 256)
             # logger.info("emb1", emb1.shape)
             c_vector = torch.cat([emb1, emb2, emb3], dim=1).to(DEVICE, dtype=DTYPE)
-            uc_vector = c_vector.clone().to(DEVICE, dtype=DTYPE)  # ちょっとここ正しいかどうかわからない I'm not sure if this is right
+            uc_vector = c_vector.clone().to(
+                DEVICE, dtype=DTYPE
+            )  # ちょっとここ正しいかどうかわからない I'm not sure if this is right
 
             # crossattn
 
@@ -215,13 +222,22 @@ def call_text_encoder(text, text2):
                 # text_embedding = pipe.text_encoder.text_model.final_layer_norm(text_embedding)    # layer normは通さないらしい
 
             # text encoder 2
-            with torch.no_grad():
-                tokens = tokenizer2(text2).to(DEVICE)
+            # tokens = tokenizer2(text2).to(DEVICE)
+            tokens = tokenizer2(
+                text,
+                truncation=True,
+                return_length=True,
+                return_overflowing_tokens=False,
+                padding="max_length",
+                return_tensors="pt",
+            )
+            tokens = batch_encoding["input_ids"].to(DEVICE)
 
+            with torch.no_grad():
                 enc_out = text_model2(tokens, output_hidden_states=True, return_dict=True)
                 text_embedding2_penu = enc_out["hidden_states"][-2]
                 # logger.info("hidden_states2", text_embedding2_penu.shape)
-                text_embedding2_pool = enc_out["text_embeds"]   # do not support Textual Inversion
+                text_embedding2_pool = enc_out["text_embeds"]  # do not support Textual Inversion
 
             # 連結して終了 concat and finish
             text_embedding = torch.cat([text_embedding1, text_embedding2_penu], dim=2)
diff --git a/sdxl_train_textual_inversion.py b/sdxl_train_textual_inversion.py
index b9a948bb2..257d181ad 100644
--- a/sdxl_train_textual_inversion.py
+++ b/sdxl_train_textual_inversion.py
@@ -7,7 +7,6 @@
 from library.device_utils import init_ipex
 init_ipex()
 
-import open_clip
 from library import sdxl_model_util, sdxl_train_util, train_util
 
 import train_textual_inversion

From 9c4492b58a7fa84126cca5f5ff39d6731865b0b5 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sun, 24 Mar 2024 23:17:25 +0900
Subject: [PATCH 49/69] fix pytorch version 2.1.1 to 2.1.2

---
 README-ja.md | 8 ++++----
 README.md    | 8 ++++----
 2 files changed, 8 insertions(+), 8 deletions(-)

diff --git a/README-ja.md b/README-ja.md
index 1d83c44f1..f70f882d7 100644
--- a/README-ja.md
+++ b/README-ja.md
@@ -41,7 +41,7 @@ PowerShellを使う場合、venvを使えるようにするためには以下の
 
 ## Windows環境でのインストール
 
-スクリプトはPyTorch 2.1.1でテストしています。PyTorch 2.0.1、1.12.1でも動作すると思われます。
+スクリプトはPyTorch 2.1.2でテストしています。PyTorch 2.0.1、1.12.1でも動作すると思われます。
 
 （なお、python -m venv～の行で「python」とだけ表示された場合、py -m venv～のようにpythonをpyに変更してください。）
 
@@ -54,9 +54,9 @@ cd sd-scripts
 python -m venv venv
 .\venv\Scripts\activate
 
-pip install torch==2.1.1 torchvision==0.16.1 --index-url https://download.pytorch.org/whl/cu118
+pip install torch==2.1.2 torchvision==0.16.2 --index-url https://download.pytorch.org/whl/cu118
 pip install --upgrade -r requirements.txt
-pip install xformers==0.0.23 --index-url https://download.pytorch.org/whl/cu118
+pip install xformers==0.0.23.post1 --index-url https://download.pytorch.org/whl/cu118
 
 accelerate config
 ```
@@ -65,7 +65,7 @@ accelerate config
 
 注：`bitsandbytes==0.43.0`、`prodigyopt==1.0`、`lion-pytorch==0.0.6` は `requirements.txt` に含まれるようになりました。他のバージョンを使う場合は適宜インストールしてください。
 
-この例では PyTorch および xfomers は2.1.1／CUDA 11.8版をインストールします。CUDA 12.1版やPyTorch 1.12.1を使う場合は適宜書き換えください。たとえば CUDA 12.1版の場合は `pip install torch==2.1.1 torchvision==0.16.1 --index-url https://download.pytorch.org/whl/cu121` および `pip install xformers==0.0.23 --index-url https://download.pytorch.org/whl/cu121` としてください。
+この例では PyTorch および xfomers は2.1.2／CUDA 11.8版をインストールします。CUDA 12.1版やPyTorch 1.12.1を使う場合は適宜書き換えください。たとえば CUDA 12.1版の場合は `pip install torch==2.1.2 torchvision==0.16.2 --index-url https://download.pytorch.org/whl/cu121` および `pip install xformers==0.0.23.post1 --index-url https://download.pytorch.org/whl/cu121` としてください。
 
 accelerate configの質問には以下のように答えてください。（bf16で学習する場合、最後の質問にはbf16と答えてください。）
 
diff --git a/README.md b/README.md
index ef26acab8..b1b924ec2 100644
--- a/README.md
+++ b/README.md
@@ -20,7 +20,7 @@ This repository contains the scripts for:
 
 The file does not contain requirements for PyTorch. Because the version of PyTorch depends on the environment, it is not included in the file. Please install PyTorch first according to the environment. See installation instructions below.
 
-The scripts are tested with Pytorch 2.1.1. 2.0.1 and 1.12.1 is not tested but should work.
+The scripts are tested with Pytorch 2.1.2. 2.0.1 and 1.12.1 is not tested but should work.
 
 ## Links to usage documentation
 
@@ -64,9 +64,9 @@ cd sd-scripts
 python -m venv venv
 .\venv\Scripts\activate
 
-pip install torch==2.1.1 torchvision==0.16.1 --index-url https://download.pytorch.org/whl/cu118
+pip install torch==2.1.2 torchvision==0.16.2 --index-url https://download.pytorch.org/whl/cu118
 pip install --upgrade -r requirements.txt
-pip install xformers==0.0.23 --index-url https://download.pytorch.org/whl/cu118
+pip install xformers==0.0.23.post1 --index-url https://download.pytorch.org/whl/cu118
 
 accelerate config
 ```
@@ -75,7 +75,7 @@ If `python -m venv` shows only `python`, change `python` to `py`.
 
 __Note:__ Now `bitsandbytes==0.43.0`, `prodigyopt==1.0` and `lion-pytorch==0.0.6` are included in the requirements.txt. If you'd like to use the another version, please install it manually.
 
-This installation is for CUDA 11.8. If you use a different version of CUDA, please install the appropriate version of PyTorch and xformers. For example, if you use CUDA 12, please install `pip install torch==2.1.1 torchvision==0.16.1 --index-url https://download.pytorch.org/whl/cu121` and `pip install xformers==0.0.23 --index-url https://download.pytorch.org/whl/cu121`.
+This installation is for CUDA 11.8. If you use a different version of CUDA, please install the appropriate version of PyTorch and xformers. For example, if you use CUDA 12, please install `pip install torch==2.1.2 torchvision==0.16.2 --index-url https://download.pytorch.org/whl/cu121` and `pip install xformers==0.0.23.post1 --index-url https://download.pytorch.org/whl/cu121`.
 
 <!-- 
 cp .\bitsandbytes_windows\*.dll .\venv\Lib\site-packages\bitsandbytes\

From a2b85316275e6a683fa70204ec192401776c2dc0 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Mon, 25 Mar 2024 22:28:46 +0900
Subject: [PATCH 50/69] make each script consistent, fix to work w/o DeepSpeed

---
 fine_tune.py                         | 7 ++++++-
 library/train_util.py                | 2 +-
 sdxl_train_control_net_lllite.py     | 7 ++++---
 sdxl_train_control_net_lllite_old.py | 7 ++++---
 train_controlnet.py                  | 4 +++-
 train_db.py                          | 2 +-
 train_network.py                     | 7 +++----
 train_textual_inversion.py           | 8 +++++---
 train_textual_inversion_XTI.py       | 4 +++-
 9 files changed, 30 insertions(+), 18 deletions(-)

diff --git a/fine_tune.py b/fine_tune.py
index e2f2c8a9a..a0350ce18 100644
--- a/fine_tune.py
+++ b/fine_tune.py
@@ -328,7 +328,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
             with accelerator.accumulate(*training_models):
                 with torch.no_grad():
                     if "latents" in batch and batch["latents"] is not None:
-                        latents = batch["latents"].to(accelerator.device)  # .to(dtype=weight_dtype)
+                        latents = batch["latents"].to(accelerator.device).to(dtype=weight_dtype)
                     else:
                         # latentに変換
                         latents = vae.encode(batch["images"].to(dtype=vae_dtype)).latent_dist.sample().to(weight_dtype)
@@ -507,6 +507,11 @@ def setup_parser() -> argparse.ArgumentParser:
         default=None,
         help="learning rate for text encoder, default is same as unet / Text Encoderの学習率、デフォルトはunetと同じ",
     )
+    parser.add_argument(
+        "--no_half_vae",
+        action="store_true",
+        help="do not use fp16/bf16 VAE in mixed precision (use float VAE) / mixed precisionでも fp16/bf16 VAEを使わずfloat VAEを使う",
+    )
 
     return parser
 
diff --git a/library/train_util.py b/library/train_util.py
index e10e38daf..6ca5d559e 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -20,7 +20,7 @@
     Tuple,
     Union,
 )
-from accelerate import Accelerator, InitProcessGroupKwargs, DistributedDataParallelKwargs
+from accelerate import Accelerator, InitProcessGroupKwargs, DistributedDataParallelKwargs, PartialState
 import glob
 import math
 import os
diff --git a/sdxl_train_control_net_lllite.py b/sdxl_train_control_net_lllite.py
index e99b4e35c..9eaaa19f2 100644
--- a/sdxl_train_control_net_lllite.py
+++ b/sdxl_train_control_net_lllite.py
@@ -22,7 +22,7 @@
 import accelerate
 from diffusers import DDPMScheduler, ControlNetModel
 from safetensors.torch import load_file
-from library import sai_model_spec, sdxl_model_util, sdxl_original_unet, sdxl_train_util
+from library import deepspeed_utils, sai_model_spec, sdxl_model_util, sdxl_original_unet, sdxl_train_util
 
 import library.model_util as model_util
 import library.train_util as train_util
@@ -394,10 +394,10 @@ def remove_model(old_ckpt_name):
             with accelerator.accumulate(unet):
                 with torch.no_grad():
                     if "latents" in batch and batch["latents"] is not None:
-                        latents = batch["latents"].to(accelerator.device)
+                        latents = batch["latents"].to(accelerator.device).to(dtype=weight_dtype)
                     else:
                         # latentに変換
-                        latents = vae.encode(batch["images"].to(dtype=vae_dtype)).latent_dist.sample()
+                        latents = vae.encode(batch["images"].to(dtype=vae_dtype)).latent_dist.sample().to(dtype=weight_dtype)
 
                         # NaNが含まれていれば警告を表示し0に置き換える
                         if torch.any(torch.isnan(latents)):
@@ -566,6 +566,7 @@ def setup_parser() -> argparse.ArgumentParser:
     train_util.add_sd_models_arguments(parser)
     train_util.add_dataset_arguments(parser, False, True, True)
     train_util.add_training_arguments(parser, False)
+    deepspeed_utils.add_deepspeed_arguments(parser)
     train_util.add_optimizer_arguments(parser)
     config_util.add_config_arguments(parser)
     custom_train_functions.add_custom_train_arguments(parser)
diff --git a/sdxl_train_control_net_lllite_old.py b/sdxl_train_control_net_lllite_old.py
index dac56eedd..e55a58896 100644
--- a/sdxl_train_control_net_lllite_old.py
+++ b/sdxl_train_control_net_lllite_old.py
@@ -18,7 +18,7 @@
 from accelerate.utils import set_seed
 from diffusers import DDPMScheduler, ControlNetModel
 from safetensors.torch import load_file
-from library import sai_model_spec, sdxl_model_util, sdxl_original_unet, sdxl_train_util
+from library import deepspeed_utils, sai_model_spec, sdxl_model_util, sdxl_original_unet, sdxl_train_util
 
 import library.model_util as model_util
 import library.train_util as train_util
@@ -361,10 +361,10 @@ def remove_model(old_ckpt_name):
             with accelerator.accumulate(network):
                 with torch.no_grad():
                     if "latents" in batch and batch["latents"] is not None:
-                        latents = batch["latents"].to(accelerator.device)
+                        latents = batch["latents"].to(accelerator.device).to(dtype=weight_dtype)
                     else:
                         # latentに変換
-                        latents = vae.encode(batch["images"].to(dtype=vae_dtype)).latent_dist.sample()
+                        latents = vae.encode(batch["images"].to(dtype=vae_dtype)).latent_dist.sample().to(dtype=weight_dtype)
 
                         # NaNが含まれていれば警告を表示し0に置き換える
                         if torch.any(torch.isnan(latents)):
@@ -534,6 +534,7 @@ def setup_parser() -> argparse.ArgumentParser:
     train_util.add_sd_models_arguments(parser)
     train_util.add_dataset_arguments(parser, False, True, True)
     train_util.add_training_arguments(parser, False)
+    deepspeed_utils.add_deepspeed_arguments(parser)
     train_util.add_optimizer_arguments(parser)
     config_util.add_config_arguments(parser)
     custom_train_functions.add_custom_train_arguments(parser)
diff --git a/train_controlnet.py b/train_controlnet.py
index e44f08853..0cb0405fd 100644
--- a/train_controlnet.py
+++ b/train_controlnet.py
@@ -11,6 +11,7 @@
 from tqdm import tqdm
 
 import torch
+from library import deepspeed_utils
 from library.device_utils import init_ipex, clean_memory_on_device
 init_ipex()
 
@@ -396,7 +397,7 @@ def remove_model(old_ckpt_name):
             with accelerator.accumulate(controlnet):
                 with torch.no_grad():
                     if "latents" in batch and batch["latents"] is not None:
-                        latents = batch["latents"].to(accelerator.device)
+                        latents = batch["latents"].to(accelerator.device).to(dtype=weight_dtype)
                     else:
                         # latentに変換
                         latents = vae.encode(batch["images"].to(dtype=weight_dtype)).latent_dist.sample()
@@ -584,6 +585,7 @@ def setup_parser() -> argparse.ArgumentParser:
     train_util.add_sd_models_arguments(parser)
     train_util.add_dataset_arguments(parser, False, True, True)
     train_util.add_training_arguments(parser, False)
+    deepspeed_utils.add_deepspeed_arguments(parser)
     train_util.add_optimizer_arguments(parser)
     config_util.add_config_arguments(parser)
     custom_train_functions.add_custom_train_arguments(parser)
diff --git a/train_db.py b/train_db.py
index e27bc7ab4..bc7ef4ed5 100644
--- a/train_db.py
+++ b/train_db.py
@@ -319,7 +319,7 @@ def train(args):
                 with torch.no_grad():
                     # latentに変換
                     if cache_latents:
-                        latents = batch["latents"].to(accelerator.device)
+                        latents = batch["latents"].to(accelerator.device).to(dtype=weight_dtype)
                     else:
                         latents = vae.encode(batch["images"].to(dtype=weight_dtype)).latent_dist.sample()
                     latents = latents * 0.18215
diff --git a/train_network.py b/train_network.py
index a74a3caf1..f2451f412 100644
--- a/train_network.py
+++ b/train_network.py
@@ -471,8 +471,7 @@ def train(self, args):
             vae.to(accelerator.device, dtype=vae_dtype)
 
         # 実験的機能：勾配も含めたfp16学習を行う　PyTorchにパッチを当ててfp16でのgrad scaleを有効にする
-        if args.full_fp16 and not args.deepspeed:
-            # During deepseed training, accelerate not handles fp16/bf16|mixed precision directly via scaler. Let deepspeed engine do.
+        if args.full_fp16:
             train_util.patch_accelerator_for_fp16_training(accelerator)
 
         # resumeする
@@ -781,11 +780,11 @@ def remove_model(old_ckpt_name):
                     on_step_start(text_encoder, unet)
 
                     if "latents" in batch and batch["latents"] is not None:
-                        latents = batch["latents"].to(accelerator.device)
+                        latents = batch["latents"].to(accelerator.device).to(dtype=weight_dtype)
                     else:
                         with torch.no_grad():
                             # latentに変換
-                            latents = vae.encode(batch["images"].to(dtype=vae_dtype)).latent_dist.sample()
+                            latents = vae.encode(batch["images"].to(dtype=vae_dtype)).latent_dist.sample().to(dtype=weight_dtype)
 
                             # NaNが含まれていれば警告を表示し0に置き換える
                             if torch.any(torch.isnan(latents)):
diff --git a/train_textual_inversion.py b/train_textual_inversion.py
index 0266bc143..72e26a450 100644
--- a/train_textual_inversion.py
+++ b/train_textual_inversion.py
@@ -8,12 +8,13 @@
 
 import torch
 from library.device_utils import init_ipex, clean_memory_on_device
+
 init_ipex()
 
 from accelerate.utils import set_seed
 from diffusers import DDPMScheduler
 from transformers import CLIPTokenizer
-from library import model_util
+from library import deepspeed_utils, model_util
 
 import library.train_util as train_util
 import library.huggingface_util as huggingface_util
@@ -558,10 +559,10 @@ def remove_model(old_ckpt_name):
                 with accelerator.accumulate(text_encoders[0]):
                     with torch.no_grad():
                         if "latents" in batch and batch["latents"] is not None:
-                            latents = batch["latents"].to(accelerator.device)
+                            latents = batch["latents"].to(accelerator.device).to(dtype=weight_dtype)
                         else:
                             # latentに変換
-                            latents = vae.encode(batch["images"].to(dtype=vae_dtype)).latent_dist.sample()
+                            latents = vae.encode(batch["images"].to(dtype=vae_dtype)).latent_dist.sample().to(dtype=weight_dtype)
                         latents = latents * self.vae_scale_factor
 
                     # Get the text embedding for conditioning
@@ -749,6 +750,7 @@ def setup_parser() -> argparse.ArgumentParser:
     train_util.add_sd_models_arguments(parser)
     train_util.add_dataset_arguments(parser, True, True, False)
     train_util.add_training_arguments(parser, True)
+    deepspeed_utils.add_deepspeed_arguments(parser)
     train_util.add_optimizer_arguments(parser)
     config_util.add_config_arguments(parser)
     custom_train_functions.add_custom_train_arguments(parser, False)
diff --git a/train_textual_inversion_XTI.py b/train_textual_inversion_XTI.py
index ad7c267eb..00b251ba9 100644
--- a/train_textual_inversion_XTI.py
+++ b/train_textual_inversion_XTI.py
@@ -8,6 +8,7 @@
 from tqdm import tqdm
 
 import torch
+from library import deepspeed_utils
 from library.device_utils import init_ipex, clean_memory_on_device
 init_ipex()
 
@@ -439,7 +440,7 @@ def remove_model(old_ckpt_name):
             with accelerator.accumulate(text_encoder):
                 with torch.no_grad():
                     if "latents" in batch and batch["latents"] is not None:
-                        latents = batch["latents"].to(accelerator.device)
+                        latents = batch["latents"].to(accelerator.device).to(dtype=weight_dtype)
                     else:
                         # latentに変換
                         latents = vae.encode(batch["images"].to(dtype=weight_dtype)).latent_dist.sample()
@@ -662,6 +663,7 @@ def setup_parser() -> argparse.ArgumentParser:
     train_util.add_sd_models_arguments(parser)
     train_util.add_dataset_arguments(parser, True, True, False)
     train_util.add_training_arguments(parser, True)
+    deepspeed_utils.add_deepspeed_arguments(parser)
     train_util.add_optimizer_arguments(parser)
     config_util.add_config_arguments(parser)
     custom_train_functions.add_custom_train_arguments(parser, False)

From 6c08e97e1fc2ae284a38a463049d9456624fd1e9 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Tue, 26 Mar 2024 20:48:08 +0900
Subject: [PATCH 51/69] update readme

---
 README.md | 107 ++++++++++++++----------------------------------------
 1 file changed, 28 insertions(+), 79 deletions(-)

diff --git a/README.md b/README.md
index d9b2ddc44..155cc048f 100644
--- a/README.md
+++ b/README.md
@@ -133,21 +133,18 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
 
 ## Change History
 
-### Masked loss
+### Mar XX, 2024 / 2024/3/XX: v0.8.6
 
-`train_network.py`, `sdxl_train_network.py` and `sdxl_train.py` now support the masked loss. `--masked_loss` option is added. 
-
-NOTE: `train_network.py` and `sdxl_train.py` are not tested yet.
-
-ControlNet dataset is used to specify the mask. The mask images should be the RGB images. The pixel value 255 in R channel is treated as the mask (the loss is calculated only for the pixels with the mask), and 0 is treated as the non-mask. See details for the dataset specification in the [LLLite documentation](./docs/train_lllite_README.md#preparing-the-dataset).
-
-
-### Working in progress
 
+- The dependent libraries are updated. Please see [Upgrade](#upgrade) and update the libraries.
+  - Especially `imagesize` is newly added, so if you cannot update immediately, please install with `pip install imagesize==1.4.1`.
+  - `bitsandbytes==0.43.0`, `prodigyopt==1.0`, `lion-pytorch==0.0.6` are included in the requirements.txt.
 - Colab seems to stop with log output. Try specifying `--console_log_simple` option in the training script to disable rich logging.
 - The `.toml` file for the dataset config is now read in UTF-8 encoding. PR [#1167](https://github.com/kohya-ss/sd-scripts/pull/1167) Thanks to Horizon1704!
 - Fixed a bug that the last subset settings are applied to all images when multiple subsets of regularization images are specified in the dataset settings. The settings for each subset are correctly applied to each image. PR [#1205](https://github.com/kohya-ss/sd-scripts/pull/1205) Thanks to feffy380!
 - `train_network.py` and `sdxl_train_network.py` are modified to record some dataset settings in the metadata of the trained model (`caption_prefix`, `caption_suffix`, `keep_tokens_separator`, `secondary_separator`, `enable_wildcard`).
+- DeepSpeed is supported. PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101)  and [#1139](https://github.com/kohya-ss/sd-scripts/pull/1139) Thanks to BootsofLagrangian! See PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101) for details.
+- The masked loss is supported in each training script. PR [#1207](https://github.com/kohya-ss/sd-scripts/pull/1207) See [Masked loss](#masked-loss) for details.
 - Some features are added to the dataset subset settings.
   - `secondary_separator` is added to specify the tag separator that is not the target of shuffling or dropping. 
     - Specify `secondary_separator=";;;"`. When you specify `secondary_separator`, the part is not shuffled or dropped. 
@@ -155,6 +152,7 @@ ControlNet dataset is used to specify the mask. The mask images should be the RG
   - `keep_tokens_separator` is updated to be used twice in the caption. When you specify `keep_tokens_separator="|||"`, the part divided by the second `|||` is not shuffled or dropped and remains at the end.
   - The existing features `caption_prefix` and `caption_suffix` can be used together. `caption_prefix` and `caption_suffix` are processed first, and then `enable_wildcard`, `keep_tokens_separator`, shuffling and dropping, and `secondary_separator` are processed in order.
   - See [Dataset config](./docs/config_README-en.md) for details.
+- The dataset with DreamBooth method supports caching image information (size, caption). PR [#1178](https://github.com/kohya-ss/sd-scripts/pull/1178) and [#1206](https://github.com/kohya-ss/sd-scripts/pull/1206) Thanks to KohakuBlueleaf! See [DreamBooth method specific options](./docs/config_README-en.md#dreambooth-specific-options) for details.
 - The support for v3 repositories is added to `tag_image_by_wd14_tagger.py` (`--onnx` option only). PR [#1192](https://github.com/kohya-ss/sd-scripts/pull/1192) Thanks to sdbds!
   - Onnx may need to be updated. Onnx is not installed by default, so please install or update it with `pip install onnx==1.15.0 onnxruntime-gpu==1.17.1` etc. Please also check the comments in `requirements.txt`.
 - The model is now saved in the subdirectory as `--repo_id` in `tag_image_by_wd14_tagger.py` . This caches multiple repo_id models. Please delete unnecessary files under `--model_dir`.
@@ -163,17 +161,22 @@ ControlNet dataset is used to specify the mask. The mask images should be the RG
 - The options `--sample_every_n_epochs` and `--sample_every_n_steps` in each training script now display a warning and ignore them when a number less than or equal to `0` is specified. Thanks to S-Del for raising the issue.
 - The [English version of the dataset settings documentation](./docs/config_README-en.md) is added. PR [#1175](https://github.com/kohya-ss/sd-scripts/pull/1175) Thanks to darkstorm2150!
 
-
+- 依存ライブラリが更新されました。[アップグレード](./README-ja.md#アップグレード) を参照しライブラリを更新してください。
+  - 特に `imagesize` が新しく追加されていますので、すぐに更新ができない場合は `pip install imagesize==1.4.1` でインストールしてください。
+  - `bitsandbytes==0.43.0`、`prodigyopt==1.0`、`lion-pytorch==0.0.6` が requirements.txt に含まれるようになりました。
 - Colab での動作時、ログ出力で停止してしまうようです。学習スクリプトに `--console_log_simple` オプションを指定し、rich のロギングを無効してお試しください。
 - データセット設定の `.toml` ファイルが UTF-8 encoding で読み込まれるようになりました。PR [#1167](https://github.com/kohya-ss/sd-scripts/pull/1167) Horizon1704 氏に感謝します。
 - データセット設定で、正則化画像のサブセットを複数指定した時、最後のサブセットの各種設定がすべてのサブセットの画像に適用される不具合が修正されました。それぞれのサブセットの設定が、それぞれの画像に正しく適用されます。PR [#1205](https://github.com/kohya-ss/sd-scripts/pull/1205) feffy380 氏に感謝します。
 - `train_network.py` および `sdxl_train_network.py` で、学習したモデルのメタデータに一部のデータセット設定が記録されるよう修正しました（`caption_prefix`、`caption_suffix`、`keep_tokens_separator`、`secondary_separator`、`enable_wildcard`）。
+- DeepSpeed がサポートされました。PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101) 、[#1139](https://github.com/kohya-ss/sd-scripts/pull/1139) BootsofLagrangian 氏に感謝します。詳細は PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101) をご覧ください。
+- 各学習スクリプトでマスクロスをサポートしました。PR [#1207](https://github.com/kohya-ss/sd-scripts/pull/1207) 詳細は [Masked loss](#masked-loss) をご覧ください。
 - データセットのサブセット設定にいくつかの機能を追加しました。
   - シャッフルの対象とならないタグ分割識別子の指定 `secondary_separator` を追加しました。`secondary_separator=";;;"` のように指定します。`secondary_separator` で区切ることで、その部分はシャッフル、drop 時にまとめて扱われます。
   - `enable_wildcard` を追加しました。`true` にするとワイルドカード記法 `{aaa|bbb|ccc}` が使えます。また複数行キャプションも有効になります。
   - `keep_tokens_separator` をキャプション内に 2 つ使えるようにしました。たとえば `keep_tokens_separator="|||"` と指定したとき、`1girl, hatsune miku, vocaloid ||| stage, mic ||| best quality, rating: general` とキャプションを指定すると、二番目の `|||` で分割された部分はシャッフル、drop されず末尾に残ります。
   - 既存の機能 `caption_prefix` と `caption_suffix` とあわせて使えます。`caption_prefix` と `caption_suffix` は一番最初に処理され、その後、ワイルドカード、`keep_tokens_separator`、シャッフルおよび drop、`secondary_separator` の順に処理されます。
   - 詳細は [データセット設定](./docs/config_README-ja.md) をご覧ください。
+- DreamBooth 方式の DataSet で画像情報（サイズ、キャプション）をキャッシュする機能が追加されました。PR [#1178](https://github.com/kohya-ss/sd-scripts/pull/1178)、[#1206](https://github.com/kohya-ss/sd-scripts/pull/1206) KohakuBlueleaf 氏に感謝します。詳細は [データセット設定](./docs/config_README-ja.md#dreambooth-方式専用のオプション) をご覧ください。
 - `tag_image_by_wd14_tagger.py` で v3 のリポジトリがサポートされました（`--onnx` 指定時のみ有効）。 PR [#1192](https://github.com/kohya-ss/sd-scripts/pull/1192) sdbds 氏に感謝します。
   - Onnx のバージョンアップが必要になるかもしれません。デフォルトでは Onnx はインストールされていませんので、`pip install onnx==1.15.0 onnxruntime-gpu==1.17.1` 等でインストール、アップデートしてください。`requirements.txt` のコメントもあわせてご確認ください。
 - `tag_image_by_wd14_tagger.py` で、モデルを`--repo_id` のサブディレクトリに保存するようにしました。これにより複数のモデルファイルがキャッシュされます。`--model_dir` 直下の不要なファイルは削除願います。
@@ -182,78 +185,24 @@ ControlNet dataset is used to specify the mask. The mask images should be the RG
 - 各学習スクリプトで `--sample_every_n_epochs` および `--sample_every_n_steps` オプションに `0` 以下の数値を指定した時、警告を表示するとともにそれらを無視するよう変更しました。問題提起していただいた S-Del 氏に感謝します。
 - データセット設定の[英語版ドキュメント](./docs/config_README-en.md) が追加されました。PR [#1175](https://github.com/kohya-ss/sd-scripts/pull/1175) darkstorm2150 氏に感謝します。
 
-
-### Mar 15, 2024 / 2024/3/15: v0.8.5
-
-- Fixed a bug that the value of timestep embedding during SDXL training was incorrect.
-  - Please update for SDXL training.
-  - The inference with the generation script is also fixed.
-  - This fix appears to resolve an issue where unintended artifacts occurred in trained models under certain conditions. 
-We would like to express our deep gratitude to Mark Saint (cacoe) from leonardo.ai, for reporting the issue and cooperating with the verification, and to gcem156 for the advice provided in identifying the part of the code that needed to be fixed.
-
-- SDXL 学習時の timestep embedding の値が誤っていたのを修正しました。
-  - SDXL の学習時にはアップデートをお願いいたします。
-  - 生成スクリプトでの推論時についてもあわせて修正しました。
-  - この修正により、特定の条件下で学習されたモデルに意図しないアーティファクトが発生する問題が解消されるようです。問題を報告いただき、また検証にご協力いただいた leonardo.ai の Mark Saint (cacoe) 氏、および修正点の特定に関するアドバイスをいただいた gcem156 氏に深く感謝いたします。
-
-### Feb 24, 2024 / 2024/2/24: v0.8.4
-
-- The log output has been improved. PR [#905](https://github.com/kohya-ss/sd-scripts/pull/905) Thanks to shirayu!
-  - The log is formatted by default. The `rich` library is required. Please see [Upgrade](#upgrade) and update the library.
-  - If `rich` is not installed, the log output will be the same as before.
-  - The following options are available in each training script:
-  - `--console_log_simple` option can be used to switch to the previous log output.
-  - `--console_log_level` option can be used to specify the log level. The default is `INFO`.
-  - `--console_log_file` option can be used to output the log to a file. The default is `None` (output to the console).
-- The sample image generation during multi-GPU training is now done with multiple GPUs. PR [#1061](https://github.com/kohya-ss/sd-scripts/pull/1061) Thanks to DKnight54!
-- The support for mps devices is improved. PR [#1054](https://github.com/kohya-ss/sd-scripts/pull/1054) Thanks to akx! If mps device exists instead of CUDA, the mps device is used automatically.
-- The `--new_conv_rank` option to specify the new rank of Conv2d is added to `networks/resize_lora.py`. PR [#1102](https://github.com/kohya-ss/sd-scripts/pull/1102) Thanks to mgz-dev!
-- An option `--highvram` to disable the optimization for environments with little VRAM is added to the training scripts. If you specify it when there is enough VRAM, the operation will be faster.
-  - Currently, only the cache part of latents is optimized.
-- The IPEX support is improved. PR [#1086](https://github.com/kohya-ss/sd-scripts/pull/1086) Thanks to Disty0!
-- Fixed a bug that `svd_merge_lora.py` crashes in some cases. PR [#1087](https://github.com/kohya-ss/sd-scripts/pull/1087) Thanks to mgz-dev!
-- DyLoRA is fixed to work with SDXL. PR [#1126](https://github.com/kohya-ss/sd-scripts/pull/1126) Thanks to tamlog06!
-- The common image generation script `gen_img.py` for SD 1/2 and SDXL is added. The basic functions are the same as the scripts for SD 1/2 and SDXL, but some new features are added.
-  - External scripts to generate prompts can be supported. It can be called with `--from_module` option. (The documentation will be added later)
-  - The normalization method after prompt weighting can be specified with `--emb_normalize_mode` option. `original` is the original method, `abs` is the normalization with the average of the absolute values, `none` is no normalization.
-- Gradual Latent Hires fix is added to each generation script. See [here](./docs/gen_img_README-ja.md#about-gradual-latent) for details.
-
-- ログ出力が改善されました。 PR [#905](https://github.com/kohya-ss/sd-scripts/pull/905) shirayu 氏に感謝します。
-  - デフォルトでログが成形されます。`rich` ライブラリが必要なため、[Upgrade](#upgrade) を参照し更新をお願いします。
-  - `rich` がインストールされていない場合は、従来のログ出力になります。
-  - 各学習スクリプトでは以下のオプションが有効です。
-  - `--console_log_simple` オプションで従来のログ出力に切り替えられます。
-  - `--console_log_level` でログレベルを指定できます。デフォルトは `INFO` です。
-  - `--console_log_file` でログファイルを出力できます。デフォルトは `None`（コンソールに出力） です。
-- 複数 GPU 学習時に学習中のサンプル画像生成を複数 GPU で行うようになりました。 PR [#1061](https://github.com/kohya-ss/sd-scripts/pull/1061) DKnight54 氏に感謝します。
-- mps デバイスのサポートが改善されました。 PR [#1054](https://github.com/kohya-ss/sd-scripts/pull/1054) akx 氏に感謝します。CUDA ではなく mps が存在する場合には自動的に mps デバイスを使用します。
-- `networks/resize_lora.py` に Conv2d の新しいランクを指定するオプション `--new_conv_rank` が追加されました。 PR [#1102](https://github.com/kohya-ss/sd-scripts/pull/1102) mgz-dev 氏に感謝します。
-- 学習スクリプトに VRAMが少ない環境向け最適化を無効にするオプション `--highvram` を追加しました。VRAM に余裕がある場合に指定すると動作が高速化されます。
-  - 現在は latents のキャッシュ部分のみ高速化されます。
-- IPEX サポートが改善されました。 PR [#1086](https://github.com/kohya-ss/sd-scripts/pull/1086) Disty0 氏に感謝します。
-- `svd_merge_lora.py` が場合によってエラーになる不具合が修正されました。 PR [#1087](https://github.com/kohya-ss/sd-scripts/pull/1087) mgz-dev 氏に感謝します。
-- DyLoRA が SDXL で動くよう修正されました。PR [#1126](https://github.com/kohya-ss/sd-scripts/pull/1126) tamlog06 氏に感謝します。
-- SD 1/2 および SDXL 共通の生成スクリプト `gen_img.py` を追加しました。基本的な機能は SD 1/2、SDXL 向けスクリプトと同じですが、いくつかの新機能が追加されています。
-  - プロンプトを動的に生成する外部スクリプトをサポートしました。 `--from_module` で呼び出せます。（ドキュメントはのちほど追加します）
-  - プロンプト重みづけ後の正規化方法を `--emb_normalize_mode` で指定できます。`original` は元の方法、`abs` は絶対値の平均値で正規化、`none` は正規化を行いません。
-- Gradual Latent Hires fix を各生成スクリプトに追加しました。詳細は [こちら](./docs/gen_img_README-ja.md#about-gradual-latent)。
-
-
-### Jan 27, 2024 / 2024/1/27: v0.8.3
-
-- Fixed a bug that the training crashes when `--fp8_base` is specified with `--save_state`. PR [#1079](https://github.com/kohya-ss/sd-scripts/pull/1079) Thanks to feffy380!
-  - `safetensors` is updated. Please see [Upgrade](#upgrade) and update the library.
-- Fixed a bug that the training crashes when `network_multiplier` is specified with multi-GPU training. PR [#1084](https://github.com/kohya-ss/sd-scripts/pull/1084) Thanks to fireicewolf!
-- Fixed a bug that the training crashes when training ControlNet-LLLite.
-
-- `--fp8_base` 指定時に `--save_state` での保存がエラーになる不具合が修正されました。 PR [#1079](https://github.com/kohya-ss/sd-scripts/pull/1079) feffy380 氏に感謝します。
-  - `safetensors` がバージョンアップされていますので、[Upgrade](#upgrade) を参照し更新をお願いします。
-- 複数 GPU での学習時に `network_multiplier` を指定するとクラッシュする不具合が修正されました。 PR [#1084](https://github.com/kohya-ss/sd-scripts/pull/1084) fireicewolf 氏に感謝します。
-- ControlNet-LLLite の学習がエラーになる不具合を修正しました。 
-
 Please read [Releases](https://github.com/kohya-ss/sd-scripts/releases) for recent updates.
 最近の更新情報は [Release](https://github.com/kohya-ss/sd-scripts/releases) をご覧ください。
 
+#### Masked loss
+
+The masked loss is supported in each training script. To enable the masked loss, specify the `--masked_loss` option.
+
+The feature is not fully tested, so there may be bugs. If you find any issues, please open an Issue.
+
+ControlNet dataset is used to specify the mask. The mask images should be the RGB images. The pixel value 255 in R channel is treated as the mask (the loss is calculated only for the pixels with the mask), and 0 is treated as the non-mask. The pixel values 0-255 are converted to 0-1 (i.e., the pixel value 128 is treated as the half weight of the loss). See details for the dataset specification in the [LLLite documentation](./docs/train_lllite_README.md#preparing-the-dataset).
+
+各学習スクリプトでマスクロスをサポートしました。マスクロスを有効にするには `--masked_loss` オプションを指定してください。
+
+機能は完全にテストされていないため、不具合があるかもしれません。その場合は Issue を立てていただけると助かります。
+
+マスクの指定には ControlNet データセットを使用します。マスク画像は RGB 画像である必要があります。R チャンネルのピクセル値 255 がロス計算対象、0 がロス計算対象外になります。0-255 の値は、0-1 の範囲に変換されます（つまりピクセル値 128 の部分はロスの重みが半分になります）。データセットの詳細は [LLLite ドキュメント](./docs/train_lllite_README-ja.md#データセットの準備) をご覧ください。
+
+
 ## Additional Information
 
 ### Naming of LoRA

From 6f7e93d5cc4e4d879e1889af8061e350d86d61d1 Mon Sep 17 00:00:00 2001
From: Disty0 <semihgulec2005@gmail.com>
Date: Wed, 27 Mar 2024 03:21:13 +0300
Subject: [PATCH 52/69] Add OpenVINO and ROCm ONNX Runtime for WD14

---
 finetune/tag_images_by_wd14_tagger.py | 23 +++++++++++++++++------
 library/ipex/attention.py             | 16 ++++++++--------
 library/ipex/hijacks.py               | 14 ++++++++------
 3 files changed, 33 insertions(+), 20 deletions(-)

diff --git a/finetune/tag_images_by_wd14_tagger.py b/finetune/tag_images_by_wd14_tagger.py
index 401c6d1ec..c19ae9749 100644
--- a/finetune/tag_images_by_wd14_tagger.py
+++ b/finetune/tag_images_by_wd14_tagger.py
@@ -142,12 +142,23 @@ def main(args):
 
         del model
 
-        ort_sess = ort.InferenceSession(
-            onnx_path,
-            providers=(
-                ["CUDAExecutionProvider"] if "CUDAExecutionProvider" in ort.get_available_providers() else ["CPUExecutionProvider"]
-            ),
-        )
+        if "OpenVINOExecutionProvider" in ort.get_available_providers():
+            # requires provider options for gpu support
+            # fp16 causes nonsense outputs
+            ort_sess = ort.InferenceSession(
+                onnx_path,
+                providers=(["OpenVINOExecutionProvider"]),
+                provider_options=[{'device_type' : "GPU_FP32"}],
+            )
+        else:
+            ort_sess = ort.InferenceSession(
+                onnx_path,
+                providers=(
+                    ["CUDAExecutionProvider"] if "CUDAExecutionProvider" in ort.get_available_providers() else
+                    ["ROCMExecutionProvider"] if "ROCMExecutionProvider" in ort.get_available_providers() else
+                    ["CPUExecutionProvider"]
+                ),
+            )
     else:
         from tensorflow.keras.models import load_model
 
diff --git a/library/ipex/attention.py b/library/ipex/attention.py
index 8253c5b17..d989ad53d 100644
--- a/library/ipex/attention.py
+++ b/library/ipex/attention.py
@@ -122,15 +122,15 @@ def torch_bmm_32_bit(input, mat2, *, out=None):
                     mat2[start_idx:end_idx],
                     out=out
                 )
+        torch.xpu.synchronize(input.device)
     else:
         return original_torch_bmm(input, mat2, out=out)
-    torch.xpu.synchronize(input.device)
     return hidden_states
 
 original_scaled_dot_product_attention = torch.nn.functional.scaled_dot_product_attention
-def scaled_dot_product_attention_32_bit(query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False):
+def scaled_dot_product_attention_32_bit(query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False, **kwargs):
     if query.device.type != "xpu":
-        return original_scaled_dot_product_attention(query, key, value, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=is_causal)
+        return original_scaled_dot_product_attention(query, key, value, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=is_causal, **kwargs)
     do_split, do_split_2, do_split_3, split_slice_size, split_2_slice_size, split_3_slice_size = find_sdpa_slice_sizes(query.shape, query.element_size())
 
     # Slice SDPA
@@ -153,7 +153,7 @@ def scaled_dot_product_attention_32_bit(query, key, value, attn_mask=None, dropo
                                 key[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3],
                                 value[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3],
                                 attn_mask=attn_mask[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3] if attn_mask is not None else attn_mask,
-                                dropout_p=dropout_p, is_causal=is_causal
+                                dropout_p=dropout_p, is_causal=is_causal, **kwargs
                             )
                     else:
                         hidden_states[start_idx:end_idx, start_idx_2:end_idx_2] = original_scaled_dot_product_attention(
@@ -161,7 +161,7 @@ def scaled_dot_product_attention_32_bit(query, key, value, attn_mask=None, dropo
                             key[start_idx:end_idx, start_idx_2:end_idx_2],
                             value[start_idx:end_idx, start_idx_2:end_idx_2],
                             attn_mask=attn_mask[start_idx:end_idx, start_idx_2:end_idx_2] if attn_mask is not None else attn_mask,
-                            dropout_p=dropout_p, is_causal=is_causal
+                            dropout_p=dropout_p, is_causal=is_causal, **kwargs
                         )
             else:
                 hidden_states[start_idx:end_idx] = original_scaled_dot_product_attention(
@@ -169,9 +169,9 @@ def scaled_dot_product_attention_32_bit(query, key, value, attn_mask=None, dropo
                     key[start_idx:end_idx],
                     value[start_idx:end_idx],
                     attn_mask=attn_mask[start_idx:end_idx] if attn_mask is not None else attn_mask,
-                    dropout_p=dropout_p, is_causal=is_causal
+                    dropout_p=dropout_p, is_causal=is_causal, **kwargs
                 )
+        torch.xpu.synchronize(query.device)
     else:
-        return original_scaled_dot_product_attention(query, key, value, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=is_causal)
-    torch.xpu.synchronize(query.device)
+        return original_scaled_dot_product_attention(query, key, value, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=is_causal, **kwargs)
     return hidden_states
diff --git a/library/ipex/hijacks.py b/library/ipex/hijacks.py
index b1b9ccf0e..65089f39e 100644
--- a/library/ipex/hijacks.py
+++ b/library/ipex/hijacks.py
@@ -12,7 +12,7 @@
 class DummyDataParallel(torch.nn.Module): # pylint: disable=missing-class-docstring, unused-argument, too-few-public-methods
     def __new__(cls, module, device_ids=None, output_device=None, dim=0): # pylint: disable=unused-argument
         if isinstance(device_ids, list) and len(device_ids) > 1:
-            logger.error("IPEX backend doesn't support DataParallel on multiple XPU devices")
+            print("IPEX backend doesn't support DataParallel on multiple XPU devices")
         return module.to("xpu")
 
 def return_null_context(*args, **kwargs): # pylint: disable=unused-argument
@@ -42,7 +42,7 @@ def autocast_init(self, device_type, dtype=None, enabled=True, cache_enabled=Non
 original_interpolate = torch.nn.functional.interpolate
 @wraps(torch.nn.functional.interpolate)
 def interpolate(tensor, size=None, scale_factor=None, mode='nearest', align_corners=None, recompute_scale_factor=None, antialias=False): # pylint: disable=too-many-arguments
-    if antialias or align_corners is not None:
+    if antialias or align_corners is not None or mode == 'bicubic':
         return_device = tensor.device
         return_dtype = tensor.dtype
         return original_interpolate(tensor.to("cpu", dtype=torch.float32), size=size, scale_factor=scale_factor, mode=mode,
@@ -216,7 +216,9 @@ def torch_empty(*args, device=None, **kwargs):
 
 original_torch_randn = torch.randn
 @wraps(torch.randn)
-def torch_randn(*args, device=None, **kwargs):
+def torch_randn(*args, device=None, dtype=None, **kwargs):
+    if dtype == bytes:
+        dtype = None
     if check_device(device):
         return original_torch_randn(*args, device=return_xpu(device), **kwargs)
     else:
@@ -256,11 +258,11 @@ def torch_Generator(device=None):
 
 original_torch_load = torch.load
 @wraps(torch.load)
-def torch_load(f, map_location=None, pickle_module=None, *, weights_only=False, mmap=None, **kwargs):
+def torch_load(f, map_location=None, *args, **kwargs):
     if check_device(map_location):
-        return original_torch_load(f, map_location=return_xpu(map_location), pickle_module=pickle_module, weights_only=weights_only, mmap=mmap, **kwargs)
+        return original_torch_load(f, map_location=return_xpu(map_location), *args, **kwargs)
     else:
-        return original_torch_load(f, map_location=map_location, pickle_module=pickle_module, weights_only=weights_only, mmap=mmap, **kwargs)
+        return original_torch_load(f, map_location=map_location, *args, **kwargs)
 
 
 # Hijack Functions:

From dd9763be31805f24255ca722f30bc5f6d99c73f5 Mon Sep 17 00:00:00 2001
From: Disty0 <semihgulec2005@gmail.com>
Date: Wed, 27 Mar 2024 21:53:40 +0300
Subject: [PATCH 53/69] Rating support for WD Tagger

---
 finetune/tag_images_by_wd14_tagger.py | 33 ++++++++++++++++++---------
 1 file changed, 22 insertions(+), 11 deletions(-)

diff --git a/finetune/tag_images_by_wd14_tagger.py b/finetune/tag_images_by_wd14_tagger.py
index c19ae9749..1d49afc7f 100644
--- a/finetune/tag_images_by_wd14_tagger.py
+++ b/finetune/tag_images_by_wd14_tagger.py
@@ -174,8 +174,9 @@ def main(args):
         rows = l[1:]
     assert header[0] == "tag_id" and header[1] == "name" and header[2] == "category", f"unexpected csv format: {header}"
 
-    general_tags = [row[1] for row in rows[1:] if row[2] == "0"]
-    character_tags = [row[1] for row in rows[1:] if row[2] == "4"]
+    rating_tags = [row[1] for row in rows[0:] if row[2] == "9"]
+    general_tags = [row[1] for row in rows[0:] if row[2] == "0"]
+    character_tags = [row[1] for row in rows[0:] if row[2] == "4"]
 
     # 画像を読み込む
 
@@ -202,17 +203,13 @@ def run_batch(path_imgs):
             probs = probs.numpy()
 
         for (image_path, _), prob in zip(path_imgs, probs):
-            # 最初の4つはratingなので無視する
-            # # First 4 labels are actually ratings: pick one with argmax
-            # ratings_names = label_names[:4]
-            # rating_index = ratings_names["probs"].argmax()
-            # found_rating = ratings_names[rating_index: rating_index + 1][["name", "probs"]]
+            combined_tags = []
+            rating_tag_text = ""
+            character_tag_text = ""
+            general_tag_text = ""
 
             # それ以降はタグなのでconfidenceがthresholdより高いものを追加する
             # Everything else is tags: pick any where prediction confidence > threshold
-            combined_tags = []
-            general_tag_text = ""
-            character_tag_text = ""
             for i, p in enumerate(prob[4:]):
                 if i < len(general_tags) and p >= args.general_threshold:
                     tag_name = general_tags[i]
@@ -231,7 +228,20 @@ def run_batch(path_imgs):
                     if tag_name not in undesired_tags:
                         tag_freq[tag_name] = tag_freq.get(tag_name, 0) + 1
                         character_tag_text += caption_separator + tag_name
-                        combined_tags.append(tag_name)
+                        combined_tags.insert(0,tag_name) # insert to the beggining
+
+            #最初の4つはratingなので無視する
+            # First 4 labels are actually ratings: pick one with argmax
+            ratings_names = prob[:4]
+            rating_index = ratings_names.argmax()
+            found_rating = rating_tags[rating_index]
+            if args.remove_underscore and len(found_rating) > 3:
+                found_rating = found_rating.replace("_", " ")
+
+            if found_rating not in undesired_tags:
+                tag_freq[found_rating] = tag_freq.get(found_rating, 0) + 1
+                rating_tag_text = found_rating
+                combined_tags.insert(0,found_rating) # insert to the beggining
 
             # 先頭のカンマを取る
             if len(general_tag_text) > 0:
@@ -264,6 +274,7 @@ def run_batch(path_imgs):
                 if args.debug:
                     logger.info("")
                     logger.info(f"{image_path}:")
+                    logger.info(f"\tRating tags: {rating_tag_text}")
                     logger.info(f"\tCharacter tags: {character_tag_text}")
                     logger.info(f"\tGeneral tags: {general_tag_text}")
 

From 954731d56402a463a71c0626cb22699bc4e43c3b Mon Sep 17 00:00:00 2001
From: Disty0 <semihgulec2005@gmail.com>
Date: Wed, 27 Mar 2024 22:00:59 +0300
Subject: [PATCH 54/69] fix typo

---
 finetune/tag_images_by_wd14_tagger.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/finetune/tag_images_by_wd14_tagger.py b/finetune/tag_images_by_wd14_tagger.py
index 1d49afc7f..4003210ed 100644
--- a/finetune/tag_images_by_wd14_tagger.py
+++ b/finetune/tag_images_by_wd14_tagger.py
@@ -228,7 +228,7 @@ def run_batch(path_imgs):
                     if tag_name not in undesired_tags:
                         tag_freq[tag_name] = tag_freq.get(tag_name, 0) + 1
                         character_tag_text += caption_separator + tag_name
-                        combined_tags.insert(0,tag_name) # insert to the beggining
+                        combined_tags.insert(0,tag_name) # insert to the beginning
 
             #最初の4つはratingなので無視する
             # First 4 labels are actually ratings: pick one with argmax
@@ -241,7 +241,7 @@ def run_batch(path_imgs):
             if found_rating not in undesired_tags:
                 tag_freq[found_rating] = tag_freq.get(found_rating, 0) + 1
                 rating_tag_text = found_rating
-                combined_tags.insert(0,found_rating) # insert to the beggining
+                combined_tags.insert(0,found_rating) # insert to the beginning
 
             # 先頭のカンマを取る
             if len(general_tag_text) > 0:

From 4012fd24f684d4d371a8736d7e65bee307077e33 Mon Sep 17 00:00:00 2001
From: Disty0 <semihgulec2005@gmail.com>
Date: Thu, 28 Mar 2024 21:08:16 +0300
Subject: [PATCH 55/69] IPEX fix pin_memory

---
 library/ipex/__init__.py |  7 ++++---
 library/ipex/hijacks.py  | 17 +++++++++++++++--
 2 files changed, 19 insertions(+), 5 deletions(-)

diff --git a/library/ipex/__init__.py b/library/ipex/__init__.py
index 972a3bf63..e5aba693c 100644
--- a/library/ipex/__init__.py
+++ b/library/ipex/__init__.py
@@ -32,6 +32,7 @@ def ipex_init(): # pylint: disable=too-many-statements
             torch.cuda.FloatTensor = torch.xpu.FloatTensor
             torch.Tensor.cuda = torch.Tensor.xpu
             torch.Tensor.is_cuda = torch.Tensor.is_xpu
+            torch.nn.Module.cuda = torch.nn.Module.xpu
             torch.UntypedStorage.cuda = torch.UntypedStorage.xpu
             torch.cuda._initialization_lock = torch.xpu.lazy_init._initialization_lock
             torch.cuda._initialized = torch.xpu.lazy_init._initialized
@@ -147,9 +148,9 @@ def ipex_init(): # pylint: disable=too-many-statements
 
             # C
             torch._C._cuda_getCurrentRawStream = ipex._C._getCurrentStream
-            ipex._C._DeviceProperties.multi_processor_count = ipex._C._DeviceProperties.gpu_eu_count
-            ipex._C._DeviceProperties.major = 2023
-            ipex._C._DeviceProperties.minor = 2
+            ipex._C._DeviceProperties.multi_processor_count = ipex._C._DeviceProperties.gpu_subslice_count
+            ipex._C._DeviceProperties.major = 2024
+            ipex._C._DeviceProperties.minor = 0
 
             # Fix functions with ipex:
             torch.cuda.mem_get_info = lambda device=None: [(torch.xpu.get_device_properties(device).total_memory - torch.xpu.memory_reserved(device)), torch.xpu.get_device_properties(device).total_memory]
diff --git a/library/ipex/hijacks.py b/library/ipex/hijacks.py
index 65089f39e..d3cef8276 100644
--- a/library/ipex/hijacks.py
+++ b/library/ipex/hijacks.py
@@ -190,6 +190,16 @@ def Tensor_cuda(self, device=None, *args, **kwargs):
     else:
         return original_Tensor_cuda(self, device, *args, **kwargs)
 
+original_Tensor_pin_memory = torch.Tensor.pin_memory
+@wraps(torch.Tensor.pin_memory)
+def Tensor_pin_memory(self, device=None, *args, **kwargs):
+    if device is None:
+        device = "xpu"
+    if check_device(device):
+        return original_Tensor_pin_memory(self, return_xpu(device), *args, **kwargs)
+    else:
+        return original_Tensor_pin_memory(self, device, *args, **kwargs)
+
 original_UntypedStorage_init = torch.UntypedStorage.__init__
 @wraps(torch.UntypedStorage.__init__)
 def UntypedStorage_init(*args, device=None, **kwargs):
@@ -259,10 +269,12 @@ def torch_Generator(device=None):
 original_torch_load = torch.load
 @wraps(torch.load)
 def torch_load(f, map_location=None, *args, **kwargs):
+    if map_location is None:
+        map_location = "xpu"
     if check_device(map_location):
-        return original_torch_load(f, map_location=return_xpu(map_location), *args, **kwargs)
+        return original_torch_load(f, *args, map_location=return_xpu(map_location), **kwargs)
     else:
-        return original_torch_load(f, map_location=map_location, *args, **kwargs)
+        return original_torch_load(f, *args, map_location=map_location, **kwargs)
 
 
 # Hijack Functions:
@@ -270,6 +282,7 @@ def ipex_hijacks():
     torch.tensor = torch_tensor
     torch.Tensor.to = Tensor_to
     torch.Tensor.cuda = Tensor_cuda
+    torch.Tensor.pin_memory = Tensor_pin_memory
     torch.UntypedStorage.__init__ = UntypedStorage_init
     torch.UntypedStorage.cuda = UntypedStorage_cuda
     torch.empty = torch_empty

From bc586ce190e1e85adcd7d9734636fac068bc929e Mon Sep 17 00:00:00 2001
From: Disty0 <semihgulec2005@gmail.com>
Date: Fri, 29 Mar 2024 13:56:42 +0300
Subject: [PATCH 56/69] Add --use_rating_tags and --character_tags_first for WD
 Tagger

---
 finetune/tag_images_by_wd14_tagger.py | 58 ++++++++++++++++++---------
 1 file changed, 38 insertions(+), 20 deletions(-)

diff --git a/finetune/tag_images_by_wd14_tagger.py b/finetune/tag_images_by_wd14_tagger.py
index 4003210ed..16a26179d 100644
--- a/finetune/tag_images_by_wd14_tagger.py
+++ b/finetune/tag_images_by_wd14_tagger.py
@@ -130,10 +130,10 @@ def main(args):
         input_name = model.graph.input[0].name
         try:
             batch_size = model.graph.input[0].type.tensor_type.shape.dim[0].dim_value
-        except:
+        except Exception:
             batch_size = model.graph.input[0].type.tensor_type.shape.dim[0].dim_param
 
-        if args.batch_size != batch_size and type(batch_size) != str and batch_size > 0:
+        if args.batch_size != batch_size and not isinstance(batch_size, str) and batch_size > 0:
             # some rebatch model may use 'N' as dynamic axes
             logger.warning(
                 f"Batch size {args.batch_size} doesn't match onnx model batch size {batch_size}, use model batch size {batch_size}"
@@ -169,9 +169,9 @@ def main(args):
 
     with open(os.path.join(model_location, CSV_FILE), "r", encoding="utf-8") as f:
         reader = csv.reader(f)
-        l = [row for row in reader]
-        header = l[0]  # tag_id,name,category,count
-        rows = l[1:]
+        line = [row for row in reader]
+        header = line[0]  # tag_id,name,category,count
+        rows = line[1:]
     assert header[0] == "tag_id" and header[1] == "name" and header[2] == "category", f"unexpected csv format: {header}"
 
     rating_tags = [row[1] for row in rows[0:] if row[2] == "9"]
@@ -228,20 +228,24 @@ def run_batch(path_imgs):
                     if tag_name not in undesired_tags:
                         tag_freq[tag_name] = tag_freq.get(tag_name, 0) + 1
                         character_tag_text += caption_separator + tag_name
-                        combined_tags.insert(0,tag_name) # insert to the beginning
+                        if args.character_tags_first: # insert to the beginning
+                            combined_tags.insert(0,tag_name)
+                        else:
+                            combined_tags.append(tag_name)
 
             #最初の4つはratingなので無視する
             # First 4 labels are actually ratings: pick one with argmax
-            ratings_names = prob[:4]
-            rating_index = ratings_names.argmax()
-            found_rating = rating_tags[rating_index]
-            if args.remove_underscore and len(found_rating) > 3:
-                found_rating = found_rating.replace("_", " ")
-
-            if found_rating not in undesired_tags:
-                tag_freq[found_rating] = tag_freq.get(found_rating, 0) + 1
-                rating_tag_text = found_rating
-                combined_tags.insert(0,found_rating) # insert to the beginning
+            if args.use_rating_tags:
+                ratings_names = prob[:4]
+                rating_index = ratings_names.argmax()
+                found_rating = rating_tags[rating_index]
+                if args.remove_underscore and len(found_rating) > 3:
+                    found_rating = found_rating.replace("_", " ")
+
+                if found_rating not in undesired_tags:
+                    tag_freq[found_rating] = tag_freq.get(found_rating, 0) + 1
+                    rating_tag_text = found_rating
+                    combined_tags.insert(0,found_rating) # insert to the beginning
 
             # 先頭のカンマを取る
             if len(general_tag_text) > 0:
@@ -332,7 +336,9 @@ def run_batch(path_imgs):
 
 def setup_parser() -> argparse.ArgumentParser:
     parser = argparse.ArgumentParser()
-    parser.add_argument("train_data_dir", type=str, help="directory for train images / 学習画像データのディレクトリ")
+    parser.add_argument(
+        "train_data_dir", type=str, help="directory for train images / 学習画像データのディレクトリ"
+    )
     parser.add_argument(
         "--repo_id",
         type=str,
@@ -350,7 +356,9 @@ def setup_parser() -> argparse.ArgumentParser:
         action="store_true",
         help="force downloading wd14 tagger models / wd14 taggerのモデルを再ダウンロードします",
     )
-    parser.add_argument("--batch_size", type=int, default=1, help="batch size in inference / 推論時のバッチサイズ")
+    parser.add_argument(
+        "--batch_size", type=int, default=1, help="batch size in inference / 推論時のバッチサイズ"
+    )
     parser.add_argument(
         "--max_data_loader_n_workers",
         type=int,
@@ -389,7 +397,9 @@ def setup_parser() -> argparse.ArgumentParser:
         action="store_true",
         help="replace underscores with spaces in the output tags / 出力されるタグのアンダースコアをスペースに置き換える",
     )
-    parser.add_argument("--debug", action="store_true", help="debug mode")
+    parser.add_argument(
+        "--debug", action="store_true", help="debug mode"
+    )
     parser.add_argument(
         "--undesired_tags",
         type=str,
@@ -399,10 +409,18 @@ def setup_parser() -> argparse.ArgumentParser:
     parser.add_argument(
         "--frequency_tags", action="store_true", help="Show frequency of tags for images / 画像ごとのタグの出現頻度を表示する"
     )
-    parser.add_argument("--onnx", action="store_true", help="use onnx model for inference / onnxモデルを推論に使用する")
+    parser.add_argument(
+        "--onnx", action="store_true", help="use onnx model for inference / onnxモデルを推論に使用する"
+    )
     parser.add_argument(
         "--append_tags", action="store_true", help="Append captions instead of overwriting / 上書きではなくキャプションを追記する"
     )
+    parser.add_argument(
+        "--use_rating_tags", action="store_true", help="Adds rating tags as the first tag",
+    )
+    parser.add_argument(
+        "--character_tags_first", action="store_true", help="Always inserts character tags before the general tags",
+    )
     parser.add_argument(
         "--caption_separator",
         type=str,

From f1f30ab4188223081aa96329a75bc4a99672b411 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sat, 30 Mar 2024 14:57:39 +0900
Subject: [PATCH 57/69] fix to work with num_beams>1 closes #1149

---
 finetune/blip/blip.py | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/finetune/blip/blip.py b/finetune/blip/blip.py
index 7d192cb26..13b69ffd3 100644
--- a/finetune/blip/blip.py
+++ b/finetune/blip/blip.py
@@ -134,8 +134,9 @@ def forward(self, image, caption):
     def generate(self, image, sample=False, num_beams=3, max_length=30, min_length=10, top_p=0.9, repetition_penalty=1.0):
         image_embeds = self.visual_encoder(image)
 
-        if not sample:
-            image_embeds = image_embeds.repeat_interleave(num_beams,dim=0)
+        # recent version of transformers seems to do repeat_interleave automatically
+        # if not sample:
+        #     image_embeds = image_embeds.repeat_interleave(num_beams,dim=0)
             
         image_atts = torch.ones(image_embeds.size()[:-1],dtype=torch.long).to(image.device)
         model_kwargs = {"encoder_hidden_states": image_embeds, "encoder_attention_mask":image_atts}

From 434dc408f9bd08a587f525e28a57b9781d10ce3c Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sat, 30 Mar 2024 17:12:36 +0900
Subject: [PATCH 58/69] update readme

---
 README.md | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/README.md b/README.md
index 155cc048f..34b11c70a 100644
--- a/README.md
+++ b/README.md
@@ -156,6 +156,7 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
 - The support for v3 repositories is added to `tag_image_by_wd14_tagger.py` (`--onnx` option only). PR [#1192](https://github.com/kohya-ss/sd-scripts/pull/1192) Thanks to sdbds!
   - Onnx may need to be updated. Onnx is not installed by default, so please install or update it with `pip install onnx==1.15.0 onnxruntime-gpu==1.17.1` etc. Please also check the comments in `requirements.txt`.
 - The model is now saved in the subdirectory as `--repo_id` in `tag_image_by_wd14_tagger.py` . This caches multiple repo_id models. Please delete unnecessary files under `--model_dir`.
+- Fixed an error when specifying `--beam_search` and a value of 2 or more for `--num_beams` in `make_captions.py`.
 - The options `--noise_offset_random_strength` and `--ip_noise_gamma_random_strength` are added to each training script. These options can be used to vary the noise offset and ip noise gamma in the range of 0 to the specified value. PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) Thanks to KohakuBlueleaf!
 - The options `--save_state_on_train_end` are added to each training script. PR [#1168](https://github.com/kohya-ss/sd-scripts/pull/1168) Thanks to gesen2egee!
 - The options `--sample_every_n_epochs` and `--sample_every_n_steps` in each training script now display a warning and ignore them when a number less than or equal to `0` is specified. Thanks to S-Del for raising the issue.
@@ -180,6 +181,7 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
 - `tag_image_by_wd14_tagger.py` で v3 のリポジトリがサポートされました（`--onnx` 指定時のみ有効）。 PR [#1192](https://github.com/kohya-ss/sd-scripts/pull/1192) sdbds 氏に感謝します。
   - Onnx のバージョンアップが必要になるかもしれません。デフォルトでは Onnx はインストールされていませんので、`pip install onnx==1.15.0 onnxruntime-gpu==1.17.1` 等でインストール、アップデートしてください。`requirements.txt` のコメントもあわせてご確認ください。
 - `tag_image_by_wd14_tagger.py` で、モデルを`--repo_id` のサブディレクトリに保存するようにしました。これにより複数のモデルファイルがキャッシュされます。`--model_dir` 直下の不要なファイルは削除願います。
+- `make_captions.py` で `--beam_search` を指定し `--num_beams` に2以上の値を指定した時のエラーを修正しました。
 - 各学習スクリプトに、noise offset、ip noise gammaを、それぞれ 0~指定した値の範囲で変動させるオプション `--noise_offset_random_strength` および `--ip_noise_gamma_random_strength` が追加されました。 PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) KohakuBlueleaf 氏に感謝します。
 - 各学習スクリプトに、学習終了時に state を保存する `--save_state_on_train_end` オプションが追加されました。 PR [#1168](https://github.com/kohya-ss/sd-scripts/pull/1168) gesen2egee 氏に感謝します。
 - 各学習スクリプトで `--sample_every_n_epochs` および `--sample_every_n_steps` オプションに `0` 以下の数値を指定した時、警告を表示するとともにそれらを無視するよう変更しました。問題提起していただいた S-Del 氏に感謝します。

From cae5aa0a56c2407564306810fc2138d4aea804c5 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sat, 30 Mar 2024 21:48:22 +0900
Subject: [PATCH 59/69] update wd14 tagger and doc

---
 README.md                             |  16 ++++
 docs/wd14_tagger_README-en.md         |  85 +++++++++++++++++++
 docs/wd14_tagger_README-ja.md         |  85 +++++++++++++++++++
 finetune/tag_images_by_wd14_tagger.py | 116 ++++++++++++++++++++------
 4 files changed, 277 insertions(+), 25 deletions(-)
 create mode 100644 docs/wd14_tagger_README-en.md
 create mode 100644 docs/wd14_tagger_README-ja.md

diff --git a/README.md b/README.md
index 34b11c70a..71614b9c6 100644
--- a/README.md
+++ b/README.md
@@ -156,6 +156,14 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
 - The support for v3 repositories is added to `tag_image_by_wd14_tagger.py` (`--onnx` option only). PR [#1192](https://github.com/kohya-ss/sd-scripts/pull/1192) Thanks to sdbds!
   - Onnx may need to be updated. Onnx is not installed by default, so please install or update it with `pip install onnx==1.15.0 onnxruntime-gpu==1.17.1` etc. Please also check the comments in `requirements.txt`.
 - The model is now saved in the subdirectory as `--repo_id` in `tag_image_by_wd14_tagger.py` . This caches multiple repo_id models. Please delete unnecessary files under `--model_dir`.
+- Some options are added to `tag_image_by_wd14_tagger.py`.
+  - Some are added in PR [#1216](https://github.com/kohya-ss/sd-scripts/pull/1216) Thanks to Disty0!
+  - Output rating tags `--use_rating_tags` and `--use_rating_tags_as_last_tag`
+  - Output character tags first `--character_tags_first`
+  - Expand character tags and series `--character_tag_expand`
+  - Specify tags to output first `--always_first_tags`
+  - Replace tags `--tag_replacement`
+  - See [Tagging documentation](./docs/wd14_tagger_README-en.md) for details.
 - Fixed an error when specifying `--beam_search` and a value of 2 or more for `--num_beams` in `make_captions.py`.
 - The options `--noise_offset_random_strength` and `--ip_noise_gamma_random_strength` are added to each training script. These options can be used to vary the noise offset and ip noise gamma in the range of 0 to the specified value. PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) Thanks to KohakuBlueleaf!
 - The options `--save_state_on_train_end` are added to each training script. PR [#1168](https://github.com/kohya-ss/sd-scripts/pull/1168) Thanks to gesen2egee!
@@ -181,6 +189,14 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
 - `tag_image_by_wd14_tagger.py` で v3 のリポジトリがサポートされました（`--onnx` 指定時のみ有効）。 PR [#1192](https://github.com/kohya-ss/sd-scripts/pull/1192) sdbds 氏に感謝します。
   - Onnx のバージョンアップが必要になるかもしれません。デフォルトでは Onnx はインストールされていませんので、`pip install onnx==1.15.0 onnxruntime-gpu==1.17.1` 等でインストール、アップデートしてください。`requirements.txt` のコメントもあわせてご確認ください。
 - `tag_image_by_wd14_tagger.py` で、モデルを`--repo_id` のサブディレクトリに保存するようにしました。これにより複数のモデルファイルがキャッシュされます。`--model_dir` 直下の不要なファイルは削除願います。
+- `tag_image_by_wd14_tagger.py` にいくつかのオプションを追加しました。
+  - 一部は PR [#1216](https://github.com/kohya-ss/sd-scripts/pull/1216) で追加されました。Disty0 氏に感謝します。
+  - レーティングタグを出力する `--use_rating_tags` および `--use_rating_tags_as_last_tag`
+  - キャラクタタグを最初に出力する `--character_tags_first`
+  - キャラクタタグとシリーズを展開する `--character_tag_expand`
+  - 常に最初に出力するタグを指定する `--always_first_tags`
+  - タグを置換する `--tag_replacement`
+  - 詳細は [タグ付けに関するドキュメント](./docs/wd14_tagger_README-ja.md) をご覧ください。
 - `make_captions.py` で `--beam_search` を指定し `--num_beams` に2以上の値を指定した時のエラーを修正しました。
 - 各学習スクリプトに、noise offset、ip noise gammaを、それぞれ 0~指定した値の範囲で変動させるオプション `--noise_offset_random_strength` および `--ip_noise_gamma_random_strength` が追加されました。 PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) KohakuBlueleaf 氏に感謝します。
 - 各学習スクリプトに、学習終了時に state を保存する `--save_state_on_train_end` オプションが追加されました。 PR [#1168](https://github.com/kohya-ss/sd-scripts/pull/1168) gesen2egee 氏に感謝します。
diff --git a/docs/wd14_tagger_README-en.md b/docs/wd14_tagger_README-en.md
new file mode 100644
index 000000000..7ac7f2ed0
--- /dev/null
+++ b/docs/wd14_tagger_README-en.md
@@ -0,0 +1,85 @@
+# Image Tagging using WD14Tagger
+
+This document is based on the information from this github page (https://github.com/toriato/stable-diffusion-webui-wd14-tagger#mrsmilingwolfs-model-aka-waifu-diffusion-14-tagger).
+
+Using onnx for inference is recommended. Please install onnx with the following command:
+
+```powershell
+pip install onnx==1.15.0 onnxruntime-gpu==1.17.1  
+```
+
+The model weights will be automatically downloaded from Hugging Face.
+
+# Usage
+
+Run the script to perform tagging.
+
+```powershell
+python finetune/tag_images_by_wd14_tagger.py --onnx --repo_id <model repo id> --batch_size <batch size> <training data folder>
+```
+
+For example, if using the repository `SmilingWolf/wd-swinv2-tagger-v3` with a batch size of 4, and the training data is located in the parent folder `train_data`, it would be:
+
+```powershell
+python tag_images_by_wd14_tagger.py --onnx --repo_id SmilingWolf/wd-swinv2-tagger-v3 --batch_size 4 ..\train_data
+```
+
+On the first run, the model files will be automatically downloaded to the `wd14_tagger_model` folder (the folder can be changed with an option). 
+
+Tag files will be created in the same directory as the training data images, with the same filename and a `.txt` extension.
+
+![Generated tag files](https://user-images.githubusercontent.com/52813779/208910534-ea514373-1185-4b7d-9ae3-61eb50bc294e.png)
+
+![Tags and image](https://user-images.githubusercontent.com/52813779/208910599-29070c15-7639-474f-b3e4-06bd5a3df29e.png)
+
+## Example
+
+To output in the Animagine XL 3.1 format, it would be as follows (enter on a single line in practice):
+
+```
+python tag_images_by_wd14_tagger.py --onnx --repo_id SmilingWolf/wd-swinv2-tagger-v3 
+    --batch_size 4  --remove_underscore --undesired_tags "PUT,YOUR,UNDESIRED,TAGS" --recursive 
+    --use_rating_tagss_as_last_tag --character_tags_first --character_tag_expand 
+    --always_first_tags "1girl,1boy"  ..\train_data
+```
+
+## Available Repository IDs
+
+[SmilingWolf's V2 and V3 models](https://huggingface.co/SmilingWolf) are available for use. Specify them in the format like `SmilingWolf/wd-vit-tagger-v3`. The default when omitted is `SmilingWolf/wd-v1-4-convnext-tagger-v2`.
+
+# Options 
+
+## General Options
+
+- `--onnx`: Use ONNX for inference. If not specified, TensorFlow will be used. If using TensorFlow, please install TensorFlow separately. 
+- `--batch_size`: Number of images to process at once. Default is 1. Adjust according to VRAM capacity.
+- `--caption_extension`: File extension for caption files. Default is `.txt`.
+- `--max_data_loader_n_workers`: Maximum number of workers for DataLoader. Specifying a value of 1 or more will use DataLoader to speed up image loading. If unspecified, DataLoader will not be used.
+- `--thresh`: Confidence threshold for outputting tags. Default is 0.35. Lowering the value will assign more tags but accuracy will decrease. 
+- `--general_threshold`: Confidence threshold for general tags. If omitted, same as `--thresh`.
+- `--character_threshold`: Confidence threshold for character tags. If omitted, same as `--thresh`.
+- `--recursive`: If specified, subfolders within the specified folder will also be processed recursively.
+- `--append_tags`: Append tags to existing tag files.
+- `--frequency_tags`: Output tag frequencies.  
+- `--debug`: Debug mode. Outputs debug information if specified.
+
+## Model Download
+
+- `--model_dir`: Folder to save model files. Default is `wd14_tagger_model`.  
+- `--force_download`: Re-download model files if specified.
+
+## Tag Editing
+
+- `--remove_underscore`: Remove underscores from output tags.
+- `--undesired_tags`: Specify tags not to output. Multiple tags can be specified, separated by commas. For example, `black eyes,black hair`.
+- `--use_rating_tags`: Output rating tags at the beginning of the tags.
+- `--use_rating_tags_as_last_tag`: Add rating tags at the end of the tags.
+- `--character_tags_first`: Output character tags first.
+- `--character_tag_expand`: Expand character tag series names. For example, split the tag `chara_name_(series)` into `chara_name, series`.  
+- `--always_first_tags`: Specify tags to always output first when a certain tag appears in an image. Multiple tags can be specified, separated by commas. For example, `1girl,1boy`.
+- `--caption_separator`: Separate tags with this string in the output file. Default is `, `.
+- `--tag_replacement`: Perform tag replacement. Specify in the format `tag1,tag2;tag3,tag4`. 
+
+When specifying `remove_underscore`, specify `undesired_tags`, `always_first_tags`, and `tag_replacement` without including underscores.
+
+When specifying `caption_separator`, separate `undesired_tags` and `always_first_tags` with `caption_separator`. Always separate `tag_replacement` with `,`.
diff --git a/docs/wd14_tagger_README-ja.md b/docs/wd14_tagger_README-ja.md
new file mode 100644
index 000000000..c75b77cbe
--- /dev/null
+++ b/docs/wd14_tagger_README-ja.md
@@ -0,0 +1,85 @@
+# WD14Taggerによるタグ付け
+
+こちらのgithubページ（https://github.com/toriato/stable-diffusion-webui-wd14-tagger#mrsmilingwolfs-model-aka-waifu-diffusion-14-tagger ）の情報を参考にさせていただきました。
+
+onnx を用いた推論を推奨します。以下のコマンドで onnx をインストールしてください。
+
+```powershell
+pip install onnx==1.15.0 onnxruntime-gpu==1.17.1
+```
+
+モデルの重みはHugging Faceから自動的にダウンロードしてきます。
+
+# 使い方
+
+スクリプトを実行してタグ付けを行います。
+```
+python fintune/tag_images_by_wd14_tagger.py --onnx --repo_id <モデルのrepo id> --batch_size <バッチサイズ> <教師データフォルダ>
+```
+
+レポジトリに `SmilingWolf/wd-swinv2-tagger-v3` を使用し、バッチサイズを4にして、教師データを親フォルダの `train_data`に置いた場合、以下のようになります。
+
+```
+python tag_images_by_wd14_tagger.py --onnx --repo_id SmilingWolf/wd-swinv2-tagger-v3 --batch_size 4 ..\train_data
+```
+
+初回起動時にはモデルファイルが `wd14_tagger_model` フォルダに自動的にダウンロードされます（フォルダはオプションで変えられます）。
+
+タグファイルが教師データ画像と同じディレクトリに、同じファイル名、拡張子.txtで作成されます。
+
+![生成されたタグファイル](https://user-images.githubusercontent.com/52813779/208910534-ea514373-1185-4b7d-9ae3-61eb50bc294e.png)
+
+![タグと画像](https://user-images.githubusercontent.com/52813779/208910599-29070c15-7639-474f-b3e4-06bd5a3df29e.png)
+
+## 記述例
+
+Animagine XL 3.1 方式で出力する場合、以下のようになります（実際には 1 行で入力してください）。
+
+```
+python tag_images_by_wd14_tagger.py --onnx --repo_id SmilingWolf/wd-swinv2-tagger-v3 
+    --batch_size 4  --remove_underscore --undesired_tags "PUT,YOUR,UNDESIRED,TAGS" --recursive 
+    --use_rating_tagss_as_last_tag --character_tags_first --character_tag_expand 
+    --always_first_tags "1girl,1boy"  ..\train_data
+```
+
+## 使用可能なリポジトリID
+
+[SmilingWolf 氏の V2、V3 のモデル](https://huggingface.co/SmilingWolf)が使用可能です。`SmilingWolf/wd-vit-tagger-v3` のように指定してください。省略時のデフォルトは `SmilingWolf/wd-v1-4-convnext-tagger-v2` です。
+
+# オプション
+
+## 一般オプション
+
+- `--onnx` : ONNX を使用して推論します。指定しない場合は TensorFlow を使用します。TensorFlow 使用時は別途 TensorFlow をインストールしてください。
+- `--batch_size` : 一度に処理する画像の数。デフォルトは1です。VRAMの容量に応じて増減してください。
+- `--caption_extension` : キャプションファイルの拡張子。デフォルトは `.txt` です。
+- `--max_data_loader_n_workers` : DataLoader の最大ワーカー数です。このオプションに 1 以上の数値を指定すると、DataLoader を用いて画像読み込みを高速化します。未指定時は DataLoader を用いません。
+- `--thresh` : 出力するタグの信頼度の閾値。デフォルトは0.35です。値を下げるとより多くのタグが付与されますが、精度は下がります。
+- `--general_threshold` : 一般タグの信頼度の閾値。省略時は `--thresh` と同じです。
+- `--character_threshold` : キャラクタータグの信頼度の閾値。省略時は `--thresh` と同じです。
+- `--recursive` : 指定すると、指定したフォルダ内のサブフォルダも再帰的に処理します。
+- `--append_tags` : 既存のタグファイルにタグを追加します。
+- `--frequency_tags` : タグの頻度を出力します。
+- `--debug` : デバッグモード。指定するとデバッグ情報を出力します。
+
+## モデルのダウンロード
+
+- `--model_dir` : モデルファイルの保存先フォルダ。デフォルトは `wd14_tagger_model` です。
+- `--force_download` : 指定するとモデルファイルを再ダウンロードします。
+
+## タグ編集関連
+
+- `--remove_underscore` : 出力するタグからアンダースコアを削除します。
+- `--undesired_tags` : 出力しないタグを指定します。カンマ区切りで複数指定できます。たとえば `black eyes,black hair` のように指定します。
+- `--use_rating_tags` : タグの最初にレーティングタグを出力します。
+- `--use_rating_tags_as_last_tag` : タグの最後にレーティングタグを追加します。
+- `--character_tags_first` : キャラクタータグを最初に出力します。
+- `--character_tag_expand` : キャラクタータグのシリーズ名を展開します。たとえば `chara_name_(series)` のタグを `chara_name, series` に分割します。
+- `--always_first_tags` : あるタグが画像に出力されたとき、そのタグを最初に出力するタグを指定します。カンマ区切りで複数指定できます。たとえば `1girl,1boy` のように指定します。
+- `--caption_separator` : 出力するファイルでタグをこの文字列で区切ります。デフォルトは `, ` です。
+- `--tag_replacement` : タグの置換を行います。`tag1,tag2;tag3,tag4` のように指定します。
+
+`remove_underscore` 指定時は、`undesired_tags`、`always_first_tags`、`tag_replacement` はアンダースコアを含めずに指定してください。
+
+`caption_separator` 指定時は、`undesired_tags`、`always_first_tags` は `caption_separator`  で区切ってください。`tag_replacement` は必ず `,` で区切ってください。
+
diff --git a/finetune/tag_images_by_wd14_tagger.py b/finetune/tag_images_by_wd14_tagger.py
index 16a26179d..a327bbd61 100644
--- a/finetune/tag_images_by_wd14_tagger.py
+++ b/finetune/tag_images_by_wd14_tagger.py
@@ -62,12 +62,12 @@ def __getitem__(self, idx):
         try:
             image = Image.open(img_path).convert("RGB")
             image = preprocess_image(image)
-            tensor = torch.tensor(image)
+            # tensor = torch.tensor(image) # これ Tensor に変換する必要ないな……(;･∀･)
         except Exception as e:
             logger.error(f"Could not load image path / 画像を読み込めません: {img_path}, error: {e}")
             return None
 
-        return (tensor, img_path)
+        return (image, img_path)
 
 
 def collate_fn_remove_corrupted(batch):
@@ -110,7 +110,7 @@ def main(args):
     else:
         logger.info("using existing wd14 tagger model")
 
-    # 画像を読み込む
+    # モデルを読み込む
     if args.onnx:
         import torch
         import onnx
@@ -178,8 +178,43 @@ def main(args):
     general_tags = [row[1] for row in rows[0:] if row[2] == "0"]
     character_tags = [row[1] for row in rows[0:] if row[2] == "4"]
 
-    # 画像を読み込む
+    # preprocess tags in advance
+    if args.character_tag_expand:
+        for i, tag in enumerate(character_tags):
+            if tag.endswith(")"):
+                # chara_name_(series) -> chara_name, series
+                # chara_name_(costume)_(series) -> chara_name_(costume), series
+                tags = tag.split("(")
+                character_tag = "(".join(tags[:-1])
+                if character_tag.endswith("_"):
+                    character_tag = character_tag[:-1]
+                series_tag = tags[-1].replace(")", "")
+                character_tags[i] = character_tag + args.caption_separator + series_tag
+
+    if args.remove_underscore:
+        rating_tags = [tag.replace("_", " ") if len(tag) > 3 else tag for tag in rating_tags]
+        general_tags = [tag.replace("_", " ") if len(tag) > 3 else tag for tag in general_tags]
+        character_tags = [tag.replace("_", " ") if len(tag) > 3 else tag for tag in character_tags]
+
+    if args.tag_replacement is not None:
+        # escape , and ; in tag_replacement: wd14 tag names may contain , and ;
+        escaped_tag_replacements = args.tag_replacement.replace("\\,", "@@@@").replace("\\;", "####")
+        tag_replacements = escaped_tag_replacements.split(";")
+        for tag_replacement in tag_replacements:
+            tags = tag_replacement.split(",")  # source, target
+            assert len(tags) == 2, f"tag replacement must be in the format of `source,target` / タグの置換は `置換元,置換先` の形式で指定してください: {args.tag_replacement}"
+
+            source, target = [tag.replace("@@@@", ",").replace("####", ";") for tag in tags]
+            logger.info(f"replacing tag: {source} -> {target}")
+
+            if source in general_tags:
+                general_tags[general_tags.index(source)] = target
+            elif source in character_tags:
+                character_tags[character_tags.index(source)] = target
+            elif source in rating_tags:
+                rating_tags[rating_tags.index(source)] = target
 
+    # 画像を読み込む
     train_data_dir_path = Path(args.train_data_dir)
     image_paths = train_util.glob_images_pathlib(train_data_dir_path, args.recursive)
     logger.info(f"found {len(image_paths)} images.")
@@ -188,7 +223,12 @@ def main(args):
 
     caption_separator = args.caption_separator
     stripped_caption_separator = caption_separator.strip()
-    undesired_tags = set(args.undesired_tags.split(stripped_caption_separator))
+    undesired_tags = args.undesired_tags.split(stripped_caption_separator)
+    undesired_tags = set([tag.strip() for tag in undesired_tags if tag.strip() != ""])
+
+    always_first_tags = None
+    if args.always_first_tags is not None:
+        always_first_tags = [tag for tag in args.always_first_tags.split(stripped_caption_separator) if tag.strip() != ""]
 
     def run_batch(path_imgs):
         imgs = np.array([im for _, im in path_imgs])
@@ -208,13 +248,11 @@ def run_batch(path_imgs):
             character_tag_text = ""
             general_tag_text = ""
 
-            # それ以降はタグなのでconfidenceがthresholdより高いものを追加する
-            # Everything else is tags: pick any where prediction confidence > threshold
+            # 最初の4つ以降はタグなのでconfidenceがthreshold以上のものを追加する
+            # First 4 labels are ratings, the rest are tags: pick any where prediction confidence >= threshold
             for i, p in enumerate(prob[4:]):
                 if i < len(general_tags) and p >= args.general_threshold:
                     tag_name = general_tags[i]
-                    if args.remove_underscore and len(tag_name) > 3:  # ignore emoji tags like >_< and ^_^
-                        tag_name = tag_name.replace("_", " ")
 
                     if tag_name not in undesired_tags:
                         tag_freq[tag_name] = tag_freq.get(tag_name, 0) + 1
@@ -222,30 +260,37 @@ def run_batch(path_imgs):
                         combined_tags.append(tag_name)
                 elif i >= len(general_tags) and p >= args.character_threshold:
                     tag_name = character_tags[i - len(general_tags)]
-                    if args.remove_underscore and len(tag_name) > 3:
-                        tag_name = tag_name.replace("_", " ")
 
                     if tag_name not in undesired_tags:
                         tag_freq[tag_name] = tag_freq.get(tag_name, 0) + 1
                         character_tag_text += caption_separator + tag_name
                         if args.character_tags_first: # insert to the beginning
-                            combined_tags.insert(0,tag_name)
+                            combined_tags.insert(0, tag_name)
                         else:
                             combined_tags.append(tag_name)
 
-            #最初の4つはratingなので無視する
+            # 最初の4つはratingなのでargmaxで選ぶ
             # First 4 labels are actually ratings: pick one with argmax
-            if args.use_rating_tags:
-                ratings_names = prob[:4]
-                rating_index = ratings_names.argmax()
+            if args.use_rating_tags or args.use_rating_tags_as_last_tag:
+                ratings_probs = prob[:4]
+                rating_index = ratings_probs.argmax()
                 found_rating = rating_tags[rating_index]
-                if args.remove_underscore and len(found_rating) > 3:
-                    found_rating = found_rating.replace("_", " ")
 
                 if found_rating not in undesired_tags:
                     tag_freq[found_rating] = tag_freq.get(found_rating, 0) + 1
                     rating_tag_text = found_rating
-                    combined_tags.insert(0,found_rating) # insert to the beginning
+                    if args.use_rating_tags:
+                        combined_tags.insert(0, found_rating) # insert to the beginning
+                    else:
+                        combined_tags.append(found_rating)
+
+            # 一番最初に置くタグを指定する
+            # Always put some tags at the beginning
+            if always_first_tags is not None:
+                for tag in always_first_tags:
+                    if tag in combined_tags:
+                        combined_tags.remove(tag)
+                        combined_tags.insert(0, tag)
 
             # 先頭のカンマを取る
             if len(general_tag_text) > 0:
@@ -303,9 +348,7 @@ def run_batch(path_imgs):
                 continue
 
             image, image_path = data
-            if image is not None:
-                image = image.detach().numpy()
-            else:
+            if image is None:
                 try:
                     image = Image.open(image_path)
                     if image.mode != "RGB":
@@ -407,7 +450,7 @@ def setup_parser() -> argparse.ArgumentParser:
         help="comma-separated list of undesired tags to remove from the output / 出力から除外したいタグのカンマ区切りのリスト",
     )
     parser.add_argument(
-        "--frequency_tags", action="store_true", help="Show frequency of tags for images / 画像ごとのタグの出現頻度を表示する"
+        "--frequency_tags", action="store_true", help="Show frequency of tags for images / タグの出現頻度を表示する"
     )
     parser.add_argument(
         "--onnx", action="store_true", help="use onnx model for inference / onnxモデルを推論に使用する"
@@ -416,10 +459,20 @@ def setup_parser() -> argparse.ArgumentParser:
         "--append_tags", action="store_true", help="Append captions instead of overwriting / 上書きではなくキャプションを追記する"
     )
     parser.add_argument(
-        "--use_rating_tags", action="store_true", help="Adds rating tags as the first tag",
+        "--use_rating_tags", action="store_true", help="Adds rating tags as the first tag / レーティングタグを最初のタグとして追加する",
     )
     parser.add_argument(
-        "--character_tags_first", action="store_true", help="Always inserts character tags before the general tags",
+        "--use_rating_tags_as_last_tag", action="store_true", help="Adds rating tags as the last tag / レーティングタグを最後のタグとして追加する",
+    )
+    parser.add_argument(
+        "--character_tags_first", action="store_true", help="Always inserts character tags before the general tags / characterタグを常にgeneralタグの前に出力する",
+    )
+    parser.add_argument(
+        "--always_first_tags",
+        type=str,
+        default=None,
+        help="comma-separated list of tags to always put at the beginning, e.g. `1girl,1boy`"
+        + " / 必ず先頭に置くタグのカンマ区切りリスト、例 : `1girl,1boy`",
     )
     parser.add_argument(
         "--caption_separator",
@@ -427,6 +480,19 @@ def setup_parser() -> argparse.ArgumentParser:
         default=", ",
         help="Separator for captions, include space if needed / キャプションの区切り文字、必要ならスペースを含めてください",
     )
+    parser.add_argument(
+        "--tag_replacement",
+        type=str,
+        default=None,
+        help="tag replacement in the format of `source1,target1;source2,target2; ...`. Escape `,` and `;` with `\`. e.g. `tag1,tag2;tag3,tag4`"
+        + " / タグの置換を `置換元1,置換先1;置換元2,置換先2; ...`で指定する。`\` で `,` と `;` をエスケープできる。例: `tag1,tag2;tag3,tag4`",
+    )
+    parser.add_argument(
+        "--character_tag_expand",
+        action="store_true",
+        help="expand tag tail parenthesis to another tag for character tags. `chara_name_(series)` becomes `chara_name, series`"
+        + " / キャラクタタグの末尾の括弧を別のタグに展開する。`chara_name_(series)` は `chara_name, series` になる",
+    )
 
     return parser
 

From f5323e3c4b467c09650362ecb5a509b43a1798be Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sat, 30 Mar 2024 22:10:37 +0900
Subject: [PATCH 60/69] update tagger doc

---
 docs/wd14_tagger_README-en.md | 5 ++++-
 docs/wd14_tagger_README-ja.md | 4 +++-
 2 files changed, 7 insertions(+), 2 deletions(-)

diff --git a/docs/wd14_tagger_README-en.md b/docs/wd14_tagger_README-en.md
index 7ac7f2ed0..4bab4275b 100644
--- a/docs/wd14_tagger_README-en.md
+++ b/docs/wd14_tagger_README-en.md
@@ -78,7 +78,10 @@ python tag_images_by_wd14_tagger.py --onnx --repo_id SmilingWolf/wd-swinv2-tagge
 - `--character_tag_expand`: Expand character tag series names. For example, split the tag `chara_name_(series)` into `chara_name, series`.  
 - `--always_first_tags`: Specify tags to always output first when a certain tag appears in an image. Multiple tags can be specified, separated by commas. For example, `1girl,1boy`.
 - `--caption_separator`: Separate tags with this string in the output file. Default is `, `.
-- `--tag_replacement`: Perform tag replacement. Specify in the format `tag1,tag2;tag3,tag4`. 
+- `--tag_replacement`: Perform tag replacement. Specify in the format `tag1,tag2;tag3,tag4`. For example, specify `aira tsubase,aira tsubase (uniform)` (when you want to train a specific costume).
+
+
+When using `tag_replacement`, it is applied after `character_tag_expand`.
 
 When specifying `remove_underscore`, specify `undesired_tags`, `always_first_tags`, and `tag_replacement` without including underscores.
 
diff --git a/docs/wd14_tagger_README-ja.md b/docs/wd14_tagger_README-ja.md
index c75b77cbe..3f1cf298f 100644
--- a/docs/wd14_tagger_README-ja.md
+++ b/docs/wd14_tagger_README-ja.md
@@ -77,7 +77,9 @@ python tag_images_by_wd14_tagger.py --onnx --repo_id SmilingWolf/wd-swinv2-tagge
 - `--character_tag_expand` : キャラクタータグのシリーズ名を展開します。たとえば `chara_name_(series)` のタグを `chara_name, series` に分割します。
 - `--always_first_tags` : あるタグが画像に出力されたとき、そのタグを最初に出力するタグを指定します。カンマ区切りで複数指定できます。たとえば `1girl,1boy` のように指定します。
 - `--caption_separator` : 出力するファイルでタグをこの文字列で区切ります。デフォルトは `, ` です。
-- `--tag_replacement` : タグの置換を行います。`tag1,tag2;tag3,tag4` のように指定します。
+- `--tag_replacement` : タグの置換を行います。`tag1,tag2;tag3,tag4` のように指定します。たとえば `aira tsubase,aira tsubase (uniform)` （特定の衣装を学習させたいとき）のように指定します。
+
+`tag_replacement` は `character_tag_expand` の後に適用されます。
 
 `remove_underscore` 指定時は、`undesired_tags`、`always_first_tags`、`tag_replacement` はアンダースコアを含めずに指定してください。
 

From 2c2ca9d7260d88038f85a517b786a204560497dd Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sat, 30 Mar 2024 22:55:56 +0900
Subject: [PATCH 61/69] update tagger doc

---
 docs/wd14_tagger_README-en.md | 4 ++--
 docs/wd14_tagger_README-ja.md | 3 ++-
 2 files changed, 4 insertions(+), 3 deletions(-)

diff --git a/docs/wd14_tagger_README-en.md b/docs/wd14_tagger_README-en.md
index 4bab4275b..422399a33 100644
--- a/docs/wd14_tagger_README-en.md
+++ b/docs/wd14_tagger_README-en.md
@@ -78,8 +78,8 @@ python tag_images_by_wd14_tagger.py --onnx --repo_id SmilingWolf/wd-swinv2-tagge
 - `--character_tag_expand`: Expand character tag series names. For example, split the tag `chara_name_(series)` into `chara_name, series`.  
 - `--always_first_tags`: Specify tags to always output first when a certain tag appears in an image. Multiple tags can be specified, separated by commas. For example, `1girl,1boy`.
 - `--caption_separator`: Separate tags with this string in the output file. Default is `, `.
-- `--tag_replacement`: Perform tag replacement. Specify in the format `tag1,tag2;tag3,tag4`. For example, specify `aira tsubase,aira tsubase (uniform)` (when you want to train a specific costume).
-
+- `--tag_replacement`: Perform tag replacement. Specify in the format `tag1,tag2;tag3,tag4`. If using `,` and `;`, escape them with `\`. \
+    For example, specify `aira tsubase,aira tsubase (uniform)` (when you want to train a specific costume), `aira tsubase,aira tsubase\, heir of shadows` (when the series name is not included in the tag).
 
 When using `tag_replacement`, it is applied after `character_tag_expand`.
 
diff --git a/docs/wd14_tagger_README-ja.md b/docs/wd14_tagger_README-ja.md
index 3f1cf298f..50d754ed4 100644
--- a/docs/wd14_tagger_README-ja.md
+++ b/docs/wd14_tagger_README-ja.md
@@ -77,7 +77,8 @@ python tag_images_by_wd14_tagger.py --onnx --repo_id SmilingWolf/wd-swinv2-tagge
 - `--character_tag_expand` : キャラクタータグのシリーズ名を展開します。たとえば `chara_name_(series)` のタグを `chara_name, series` に分割します。
 - `--always_first_tags` : あるタグが画像に出力されたとき、そのタグを最初に出力するタグを指定します。カンマ区切りで複数指定できます。たとえば `1girl,1boy` のように指定します。
 - `--caption_separator` : 出力するファイルでタグをこの文字列で区切ります。デフォルトは `, ` です。
-- `--tag_replacement` : タグの置換を行います。`tag1,tag2;tag3,tag4` のように指定します。たとえば `aira tsubase,aira tsubase (uniform)` （特定の衣装を学習させたいとき）のように指定します。
+- `--tag_replacement` : タグの置換を行います。`tag1,tag2;tag3,tag4` のように指定します。`,` および `;` を使う場合は `\` でエスケープしてください。\
+    たとえば `aira tsubase,aira tsubase (uniform)` （特定の衣装を学習させたいとき）、`aira tsubase,aira tsubase\, heir of shadows` （シリーズ名がタグに含まれないとき）のように指定します。
 
 `tag_replacement` は `character_tag_expand` の後に適用されます。
 

From 059ee047f3c90f37232eac9bc48c0711925d1b0b Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sat, 30 Mar 2024 23:02:24 +0900
Subject: [PATCH 62/69] fix typo

---
 docs/wd14_tagger_README-en.md | 2 +-
 docs/wd14_tagger_README-ja.md | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/docs/wd14_tagger_README-en.md b/docs/wd14_tagger_README-en.md
index 422399a33..34f448823 100644
--- a/docs/wd14_tagger_README-en.md
+++ b/docs/wd14_tagger_README-en.md
@@ -39,7 +39,7 @@ To output in the Animagine XL 3.1 format, it would be as follows (enter on a sin
 ```
 python tag_images_by_wd14_tagger.py --onnx --repo_id SmilingWolf/wd-swinv2-tagger-v3 
     --batch_size 4  --remove_underscore --undesired_tags "PUT,YOUR,UNDESIRED,TAGS" --recursive 
-    --use_rating_tagss_as_last_tag --character_tags_first --character_tag_expand 
+    --use_rating_tags_as_last_tag --character_tags_first --character_tag_expand 
     --always_first_tags "1girl,1boy"  ..\train_data
 ```
 
diff --git a/docs/wd14_tagger_README-ja.md b/docs/wd14_tagger_README-ja.md
index 50d754ed4..58e9ede95 100644
--- a/docs/wd14_tagger_README-ja.md
+++ b/docs/wd14_tagger_README-ja.md
@@ -38,7 +38,7 @@ Animagine XL 3.1 方式で出力する場合、以下のようになります（
 ```
 python tag_images_by_wd14_tagger.py --onnx --repo_id SmilingWolf/wd-swinv2-tagger-v3 
     --batch_size 4  --remove_underscore --undesired_tags "PUT,YOUR,UNDESIRED,TAGS" --recursive 
-    --use_rating_tagss_as_last_tag --character_tags_first --character_tag_expand 
+    --use_rating_tags_as_last_tag --character_tags_first --character_tag_expand 
     --always_first_tags "1girl,1boy"  ..\train_data
 ```
 

From 2258a1b753a321ed25b9ae1b7f2ceb1b24ae0736 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sun, 31 Mar 2024 15:50:35 +0900
Subject: [PATCH 63/69] add save/load hook to remove U-Net/TEs from state

---
 README.md        | 12 ++++++++++--
 train_network.py | 25 +++++++++++++++++++++++++
 2 files changed, 35 insertions(+), 2 deletions(-)

diff --git a/README.md b/README.md
index 71614b9c6..87cdf0b76 100644
--- a/README.md
+++ b/README.md
@@ -114,6 +114,10 @@ pip install --use-pep517 --upgrade -r requirements.txt
 
 Once the commands have completed successfully you should be ready to use the new version.
 
+### Upgrade PyTorch
+
+If you want to upgrade PyTorch, you can upgrade it with `pip install` command in [Windows Installation](#windows-installation) section. `xformers` is also required to be upgraded when PyTorch is upgraded.
+
 ## Credits
 
 The implementation for LoRA is based on [cloneofsimo's repo](https://github.com/cloneofsimo/lora). Thank you for great work!
@@ -137,12 +141,14 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
 
 
 - The dependent libraries are updated. Please see [Upgrade](#upgrade) and update the libraries.
-  - Especially `imagesize` is newly added, so if you cannot update immediately, please install with `pip install imagesize==1.4.1`.
+  - Especially `imagesize` is newly added, so if you cannot update the libraries immediately, please install with `pip install imagesize==1.4.1` separately.
   - `bitsandbytes==0.43.0`, `prodigyopt==1.0`, `lion-pytorch==0.0.6` are included in the requirements.txt.
+  - Also, the PyTorch version is updated to 2.1.2 (PyTorch does not need to be updated immediately). In the upgrade procedure, PyTorch is not updated, so please manually install or update torch, torchvision, xformers if necessary (see [Upgrade PyTorch](#upgrade-pytorch)).
 - Colab seems to stop with log output. Try specifying `--console_log_simple` option in the training script to disable rich logging.
 - The `.toml` file for the dataset config is now read in UTF-8 encoding. PR [#1167](https://github.com/kohya-ss/sd-scripts/pull/1167) Thanks to Horizon1704!
 - Fixed a bug that the last subset settings are applied to all images when multiple subsets of regularization images are specified in the dataset settings. The settings for each subset are correctly applied to each image. PR [#1205](https://github.com/kohya-ss/sd-scripts/pull/1205) Thanks to feffy380!
 - `train_network.py` and `sdxl_train_network.py` are modified to record some dataset settings in the metadata of the trained model (`caption_prefix`, `caption_suffix`, `keep_tokens_separator`, `secondary_separator`, `enable_wildcard`).
+- Fixed a bug that U-Net and Text Encoders are included in the state in `train_network.py` and `sdxl_train_network.py`. The saving and loading of the state are faster, the file size is smaller, and the memory usage when loading is reduced.
 - DeepSpeed is supported. PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101)  and [#1139](https://github.com/kohya-ss/sd-scripts/pull/1139) Thanks to BootsofLagrangian! See PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101) for details.
 - The masked loss is supported in each training script. PR [#1207](https://github.com/kohya-ss/sd-scripts/pull/1207) See [Masked loss](#masked-loss) for details.
 - Some features are added to the dataset subset settings.
@@ -171,12 +177,14 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
 - The [English version of the dataset settings documentation](./docs/config_README-en.md) is added. PR [#1175](https://github.com/kohya-ss/sd-scripts/pull/1175) Thanks to darkstorm2150!
 
 - 依存ライブラリが更新されました。[アップグレード](./README-ja.md#アップグレード) を参照しライブラリを更新してください。
-  - 特に `imagesize` が新しく追加されていますので、すぐに更新ができない場合は `pip install imagesize==1.4.1` でインストールしてください。
+  - 特に `imagesize` が新しく追加されていますので、すぐにライブラリの更新ができない場合は `pip install imagesize==1.4.1` で個別にインストールしてください。
   - `bitsandbytes==0.43.0`、`prodigyopt==1.0`、`lion-pytorch==0.0.6` が requirements.txt に含まれるようになりました。
+  - また PyTorch のバージョンを 2.1.2 に更新しました。PyTorch はすぐに更新する必要はありません。更新時は、アップグレードの手順では PyTorch が更新されませんので、torch、torchvision、xformers を手動でインストールしてください。
 - Colab での動作時、ログ出力で停止してしまうようです。学習スクリプトに `--console_log_simple` オプションを指定し、rich のロギングを無効してお試しください。
 - データセット設定の `.toml` ファイルが UTF-8 encoding で読み込まれるようになりました。PR [#1167](https://github.com/kohya-ss/sd-scripts/pull/1167) Horizon1704 氏に感謝します。
 - データセット設定で、正則化画像のサブセットを複数指定した時、最後のサブセットの各種設定がすべてのサブセットの画像に適用される不具合が修正されました。それぞれのサブセットの設定が、それぞれの画像に正しく適用されます。PR [#1205](https://github.com/kohya-ss/sd-scripts/pull/1205) feffy380 氏に感謝します。
 - `train_network.py` および `sdxl_train_network.py` で、学習したモデルのメタデータに一部のデータセット設定が記録されるよう修正しました（`caption_prefix`、`caption_suffix`、`keep_tokens_separator`、`secondary_separator`、`enable_wildcard`）。
+- `train_network.py` および `sdxl_train_network.py` で、state に U-Net および Text Encoder が含まれる不具合を修正しました。state の保存、読み込みが高速化され、ファイルサイズも小さくなり、また読み込み時のメモリ使用量も削減されます。
 - DeepSpeed がサポートされました。PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101) 、[#1139](https://github.com/kohya-ss/sd-scripts/pull/1139) BootsofLagrangian 氏に感謝します。詳細は PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101) をご覧ください。
 - 各学習スクリプトでマスクロスをサポートしました。PR [#1207](https://github.com/kohya-ss/sd-scripts/pull/1207) 詳細は [Masked loss](#masked-loss) をご覧ください。
 - データセットのサブセット設定にいくつかの機能を追加しました。
diff --git a/train_network.py b/train_network.py
index ed569aea6..8fe98f126 100644
--- a/train_network.py
+++ b/train_network.py
@@ -471,6 +471,31 @@ def train(self, args):
         if args.full_fp16:
             train_util.patch_accelerator_for_fp16_training(accelerator)
 
+        # before resuming make hook for saving/loading to save/load the network weights only
+        def save_model_hook(models, weights, output_dir):
+            # pop weights of other models than network to save only network weights
+            if accelerator.is_main_process:
+                remove_indices = []
+                for i,model in enumerate(models):
+                    if not isinstance(model, type(accelerator.unwrap_model(network))):
+                        remove_indices.append(i)
+                for i in reversed(remove_indices):
+                    weights.pop(i)
+                # print(f"save model hook: {len(weights)} weights will be saved")
+
+        def load_model_hook(models, input_dir):
+            # remove models except network
+            remove_indices = []
+            for i, model in enumerate(models):
+                if not isinstance(model, type(accelerator.unwrap_model(network))):
+                    remove_indices.append(i)
+            for i in reversed(remove_indices):
+                models.pop(i)
+            # print(f"load model hook: {len(models)} models will be loaded")
+
+        accelerator.register_save_state_pre_hook(save_model_hook)
+        accelerator.register_load_state_pre_hook(load_model_hook)
+
         # resumeする
         train_util.resume_from_local_or_hf_if_specified(accelerator, args)
 

From b748b48dbbaf6ddd011f032f5ede47aea094a208 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Wed, 3 Apr 2024 12:43:08 +0900
Subject: [PATCH 64/69] fix attention couple+deep shink cause error in some
 reso

---
 networks/lora.py | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/networks/lora.py b/networks/lora.py
index 948b30b0e..d1208040f 100644
--- a/networks/lora.py
+++ b/networks/lora.py
@@ -247,14 +247,13 @@ def get_mask_for_x(self, x):
             area = x.size()[1]
 
         mask = self.network.mask_dic.get(area, None)
-        if mask is None:
-            # raise ValueError(f"mask is None for resolution {area}")
+        if mask is None or len(x.size()) == 2:
             # emb_layers in SDXL doesn't have mask
             # if "emb" not in self.lora_name:
             #     print(f"mask is None for resolution {self.lora_name}, {area}, {x.size()}")
             mask_size = (1, x.size()[1]) if len(x.size()) == 2 else (1, *x.size()[1:-1], 1)
             return torch.ones(mask_size, dtype=x.dtype, device=x.device) / self.network.num_sub_prompts
-        if len(x.size()) != 4:
+        if len(x.size()) == 3:
             mask = torch.reshape(mask, (1, -1, 1))
         return mask
 

From cd587ce62cd340de32f8bf72e6eae209ce5d5580 Mon Sep 17 00:00:00 2001
From: ykume <ykumeykume@gmail.com>
Date: Fri, 5 Apr 2024 08:23:03 +0900
Subject: [PATCH 65/69] verify command line args if wandb is enabled

---
 fine_tune.py                         |  1 +
 library/train_util.py                | 56 +++++++++++++++++++++++++++-
 sdxl_train.py                        |  1 +
 sdxl_train_control_net_lllite.py     |  1 +
 sdxl_train_control_net_lllite_old.py |  1 +
 sdxl_train_network.py                |  1 +
 sdxl_train_textual_inversion.py      |  1 +
 train_controlnet.py                  |  1 +
 train_db.py                          |  1 +
 train_network.py                     |  1 +
 train_textual_inversion.py           |  1 +
 train_textual_inversion_XTI.py       |  1 +
 12 files changed, 66 insertions(+), 1 deletion(-)

diff --git a/fine_tune.py b/fine_tune.py
index a0350ce18..3c4a5a26b 100644
--- a/fine_tune.py
+++ b/fine_tune.py
@@ -520,6 +520,7 @@ def setup_parser() -> argparse.ArgumentParser:
     parser = setup_parser()
 
     args = parser.parse_args()
+    train_util.verify_command_line_training_args(args)
     args = train_util.read_config_from_file(args, parser)
 
     train(args)
diff --git a/library/train_util.py b/library/train_util.py
index 1a46f6a7d..c13bb68ee 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -1890,7 +1890,7 @@ def __init__(
                 subset.image_dir,
                 False,
                 None,
-                subset.caption_extension, 
+                subset.caption_extension,
                 subset.cache_info,
                 subset.num_repeats,
                 subset.shuffle_caption,
@@ -3358,6 +3358,60 @@ def add_masked_loss_arguments(parser: argparse.ArgumentParser):
     )
 
 
+# verify command line args for training
+def verify_command_line_training_args(args: argparse.Namespace):
+    # if wandb is enabled, the command line is exposed to the public
+    # check whether sensitive options are included in the command line arguments
+    # if so, warn or inform the user to move them to the configuration file
+    # wandbが有効な場合、コマンドラインが公開される
+    # 学習用のコマンドライン引数に敏感なオプションが含まれているかどうかを確認し、
+    # 含まれている場合は設定ファイルに移動するようにユーザーに警告または通知する
+
+    wandb_enabled = args.log_with is not None and args.log_with != "tensorboard"  # "all" or "wandb"
+    if not wandb_enabled:
+        return
+
+    sensitive_args = ["wandb_api_key", "huggingface_token"]
+    sensitive_path_args = [
+        "pretrained_model_name_or_path",
+        "vae",
+        "tokenizer_cache_dir",
+        "train_data_dir",
+        "conditioning_data_dir",
+        "reg_data_dir",
+        "output_dir",
+        "logging_dir",
+    ]
+
+    for arg in sensitive_args:
+        if getattr(args, arg, None) is not None:
+            logger.warning(
+                f"wandb is enabled, but option `{arg}` is included in the command line. Because the command line is exposed to the public, it is recommended to move it to the `.toml` file."
+                + f" / wandbが有効で、かつオプション `{arg}` がコマンドラインに含まれています。コマンドラインは公開されるため、`.toml`ファイルに移動することをお勧めします。"
+            )
+
+    # if path is absolute, it may include sensitive information
+    for arg in sensitive_path_args:
+        if getattr(args, arg, None) is not None and os.path.isabs(getattr(args, arg)):
+            logger.info(
+                f"wandb is enabled, but option `{arg}` is included in the command line and it is an absolute path. Because the command line is exposed to the public, it is recommended to move it to the `.toml` file or use relative path."
+                + f" / wandbが有効で、かつオプション `{arg}` がコマンドラインに含まれており、絶対パスです。コマンドラインは公開されるため、`.toml`ファイルに移動するか、相対パスを使用することをお勧めします。"
+            )
+
+    if getattr(args, "config_file", None) is not None:
+        logger.info(
+            f"wandb is enabled, but option `config_file` is included in the command line. Because the command line is exposed to the public, please be careful about the information included in the path."
+            + f" / wandbが有効で、かつオプション `config_file` がコマンドラインに含まれています。コマンドラインは公開されるため、パスに含まれる情報にご注意ください。"
+        )
+
+    # other sensitive options
+    if args.huggingface_repo_id is not None and args.huggingface_repo_visibility != "public":
+        logger.info(
+            f"wandb is enabled, but option huggingface_repo_id is included in the command line and huggingface_repo_visibility is not 'public'. Because the command line is exposed to the public, it is recommended to move it to the `.toml` file."
+            + f" / wandbが有効で、かつオプション huggingface_repo_id がコマンドラインに含まれており、huggingface_repo_visibility が 'public' ではありません。コマンドラインは公開されるため、`.toml`ファイルに移動することをお勧めします。"
+        )
+
+
 def verify_training_args(args: argparse.Namespace):
     r"""
     Verify training arguments. Also reflect highvram option to global variable
diff --git a/sdxl_train.py b/sdxl_train.py
index 816598e04..f6d277494 100644
--- a/sdxl_train.py
+++ b/sdxl_train.py
@@ -812,6 +812,7 @@ def setup_parser() -> argparse.ArgumentParser:
     parser = setup_parser()
 
     args = parser.parse_args()
+    train_util.verify_command_line_training_args(args)
     args = train_util.read_config_from_file(args, parser)
 
     train(args)
diff --git a/sdxl_train_control_net_lllite.py b/sdxl_train_control_net_lllite.py
index 9eaaa19f2..e880b57de 100644
--- a/sdxl_train_control_net_lllite.py
+++ b/sdxl_train_control_net_lllite.py
@@ -612,6 +612,7 @@ def setup_parser() -> argparse.ArgumentParser:
     parser = setup_parser()
 
     args = parser.parse_args()
+    train_util.verify_command_line_training_args(args)
     args = train_util.read_config_from_file(args, parser)
 
     train(args)
diff --git a/sdxl_train_control_net_lllite_old.py b/sdxl_train_control_net_lllite_old.py
index e55a58896..0ea64b824 100644
--- a/sdxl_train_control_net_lllite_old.py
+++ b/sdxl_train_control_net_lllite_old.py
@@ -580,6 +580,7 @@ def setup_parser() -> argparse.ArgumentParser:
     parser = setup_parser()
 
     args = parser.parse_args()
+    train_util.verify_command_line_training_args(args)
     args = train_util.read_config_from_file(args, parser)
 
     train(args)
diff --git a/sdxl_train_network.py b/sdxl_train_network.py
index d33239d92..83969bb1d 100644
--- a/sdxl_train_network.py
+++ b/sdxl_train_network.py
@@ -178,6 +178,7 @@ def setup_parser() -> argparse.ArgumentParser:
     parser = setup_parser()
 
     args = parser.parse_args()
+    train_util.verify_command_line_training_args(args)
     args = train_util.read_config_from_file(args, parser)
 
     trainer = SdxlNetworkTrainer()
diff --git a/sdxl_train_textual_inversion.py b/sdxl_train_textual_inversion.py
index 257d181ad..5df739e28 100644
--- a/sdxl_train_textual_inversion.py
+++ b/sdxl_train_textual_inversion.py
@@ -131,6 +131,7 @@ def setup_parser() -> argparse.ArgumentParser:
     parser = setup_parser()
 
     args = parser.parse_args()
+    train_util.verify_command_line_training_args(args)
     args = train_util.read_config_from_file(args, parser)
 
     trainer = SdxlTextualInversionTrainer()
diff --git a/train_controlnet.py b/train_controlnet.py
index 0cb0405fd..90cac0410 100644
--- a/train_controlnet.py
+++ b/train_controlnet.py
@@ -617,6 +617,7 @@ def setup_parser() -> argparse.ArgumentParser:
     parser = setup_parser()
 
     args = parser.parse_args()
+    train_util.verify_command_line_training_args(args)
     args = train_util.read_config_from_file(args, parser)
 
     train(args)
diff --git a/train_db.py b/train_db.py
index 0a152f224..c3b7339f3 100644
--- a/train_db.py
+++ b/train_db.py
@@ -523,6 +523,7 @@ def setup_parser() -> argparse.ArgumentParser:
     parser = setup_parser()
 
     args = parser.parse_args()
+    train_util.verify_command_line_training_args(args)
     args = train_util.read_config_from_file(args, parser)
 
     train(args)
diff --git a/train_network.py b/train_network.py
index 8fe98f126..fcf4cd9b6 100644
--- a/train_network.py
+++ b/train_network.py
@@ -1101,6 +1101,7 @@ def setup_parser() -> argparse.ArgumentParser:
     parser = setup_parser()
 
     args = parser.parse_args()
+    train_util.verify_command_line_training_args(args)
     args = train_util.read_config_from_file(args, parser)
 
     trainer = NetworkTrainer()
diff --git a/train_textual_inversion.py b/train_textual_inversion.py
index e7083596f..02edf9525 100644
--- a/train_textual_inversion.py
+++ b/train_textual_inversion.py
@@ -806,6 +806,7 @@ def setup_parser() -> argparse.ArgumentParser:
     parser = setup_parser()
 
     args = parser.parse_args()
+    train_util.verify_command_line_training_args(args)
     args = train_util.read_config_from_file(args, parser)
 
     trainer = TextualInversionTrainer()
diff --git a/train_textual_inversion_XTI.py b/train_textual_inversion_XTI.py
index 861d48d1d..f0723f2a7 100644
--- a/train_textual_inversion_XTI.py
+++ b/train_textual_inversion_XTI.py
@@ -714,6 +714,7 @@ def setup_parser() -> argparse.ArgumentParser:
     parser = setup_parser()
 
     args = parser.parse_args()
+    train_util.verify_command_line_training_args(args)
     args = train_util.read_config_from_file(args, parser)
 
     train(args)

From 089727b5ee40193464d5f81662b523aa7c52bee2 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sun, 7 Apr 2024 12:42:49 +0900
Subject: [PATCH 66/69] update readme

---
 README.md | 77 +++++++++++++++++++++++++++++++++++++++----------------
 1 file changed, 55 insertions(+), 22 deletions(-)

diff --git a/README.md b/README.md
index 87cdf0b76..0cecc5676 100644
--- a/README.md
+++ b/README.md
@@ -139,18 +139,34 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
 
 ### Mar XX, 2024 / 2024/3/XX: v0.8.6
 
+#### Highlights
 
 - The dependent libraries are updated. Please see [Upgrade](#upgrade) and update the libraries.
   - Especially `imagesize` is newly added, so if you cannot update the libraries immediately, please install with `pip install imagesize==1.4.1` separately.
   - `bitsandbytes==0.43.0`, `prodigyopt==1.0`, `lion-pytorch==0.0.6` are included in the requirements.txt.
   - Also, the PyTorch version is updated to 2.1.2 (PyTorch does not need to be updated immediately). In the upgrade procedure, PyTorch is not updated, so please manually install or update torch, torchvision, xformers if necessary (see [Upgrade PyTorch](#upgrade-pytorch)).
+- When logging to wandb is enabled, the entire command line is exposed. Therefore, it is recommended to write the API key of wandb and the token of HuggingFace in the configuration file (`.toml`). Thanks to bghira for raising the issue.
+  - A warning is displayed at the start of training if such information is included in the command line.
+  - Also, if there is an absolute path, the path may be exposed, so it is recommended to specify a relative path or write it in the configuration file. In such cases, an INFO log is displayed.
+  - See [#1123](https://github.com/kohya-ss/sd-scripts/pull/1123) and PR [#1240](https://github.com/kohya-ss/sd-scripts/pull/1240) for details.
 - Colab seems to stop with log output. Try specifying `--console_log_simple` option in the training script to disable rich logging.
-- The `.toml` file for the dataset config is now read in UTF-8 encoding. PR [#1167](https://github.com/kohya-ss/sd-scripts/pull/1167) Thanks to Horizon1704!
-- Fixed a bug that the last subset settings are applied to all images when multiple subsets of regularization images are specified in the dataset settings. The settings for each subset are correctly applied to each image. PR [#1205](https://github.com/kohya-ss/sd-scripts/pull/1205) Thanks to feffy380!
+- Other improvements include the addition of masked loss, DeepSpeed support, dataset settings improvements, and image tagging improvements. See below for details.
+
+#### Training scripts
+
 - `train_network.py` and `sdxl_train_network.py` are modified to record some dataset settings in the metadata of the trained model (`caption_prefix`, `caption_suffix`, `keep_tokens_separator`, `secondary_separator`, `enable_wildcard`).
 - Fixed a bug that U-Net and Text Encoders are included in the state in `train_network.py` and `sdxl_train_network.py`. The saving and loading of the state are faster, the file size is smaller, and the memory usage when loading is reduced.
 - DeepSpeed is supported. PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101)  and [#1139](https://github.com/kohya-ss/sd-scripts/pull/1139) Thanks to BootsofLagrangian! See PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101) for details.
 - The masked loss is supported in each training script. PR [#1207](https://github.com/kohya-ss/sd-scripts/pull/1207) See [Masked loss](#masked-loss) for details.
+- The options `--noise_offset_random_strength` and `--ip_noise_gamma_random_strength` are added to each training script. These options can be used to vary the noise offset and ip noise gamma in the range of 0 to the specified value. PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) Thanks to KohakuBlueleaf!
+- The options `--save_state_on_train_end` are added to each training script. PR [#1168](https://github.com/kohya-ss/sd-scripts/pull/1168) Thanks to gesen2egee!
+- The options `--sample_every_n_epochs` and `--sample_every_n_steps` in each training script now display a warning and ignore them when a number less than or equal to `0` is specified. Thanks to S-Del for raising the issue.
+
+#### Dataset settings
+
+- The [English version of the dataset settings documentation](./docs/config_README-en.md) is added. PR [#1175](https://github.com/kohya-ss/sd-scripts/pull/1175) Thanks to darkstorm2150!
+- The `.toml` file for the dataset config is now read in UTF-8 encoding. PR [#1167](https://github.com/kohya-ss/sd-scripts/pull/1167) Thanks to Horizon1704!
+- Fixed a bug that the last subset settings are applied to all images when multiple subsets of regularization images are specified in the dataset settings. The settings for each subset are correctly applied to each image. PR [#1205](https://github.com/kohya-ss/sd-scripts/pull/1205) Thanks to feffy380!
 - Some features are added to the dataset subset settings.
   - `secondary_separator` is added to specify the tag separator that is not the target of shuffling or dropping. 
     - Specify `secondary_separator=";;;"`. When you specify `secondary_separator`, the part is not shuffled or dropped. 
@@ -159,6 +175,9 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
   - The existing features `caption_prefix` and `caption_suffix` can be used together. `caption_prefix` and `caption_suffix` are processed first, and then `enable_wildcard`, `keep_tokens_separator`, shuffling and dropping, and `secondary_separator` are processed in order.
   - See [Dataset config](./docs/config_README-en.md) for details.
 - The dataset with DreamBooth method supports caching image information (size, caption). PR [#1178](https://github.com/kohya-ss/sd-scripts/pull/1178) and [#1206](https://github.com/kohya-ss/sd-scripts/pull/1206) Thanks to KohakuBlueleaf! See [DreamBooth method specific options](./docs/config_README-en.md#dreambooth-specific-options) for details.
+
+#### Image tagging
+
 - The support for v3 repositories is added to `tag_image_by_wd14_tagger.py` (`--onnx` option only). PR [#1192](https://github.com/kohya-ss/sd-scripts/pull/1192) Thanks to sdbds!
   - Onnx may need to be updated. Onnx is not installed by default, so please install or update it with `pip install onnx==1.15.0 onnxruntime-gpu==1.17.1` etc. Please also check the comments in `requirements.txt`.
 - The model is now saved in the subdirectory as `--repo_id` in `tag_image_by_wd14_tagger.py` . This caches multiple repo_id models. Please delete unnecessary files under `--model_dir`.
@@ -171,22 +190,43 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
   - Replace tags `--tag_replacement`
   - See [Tagging documentation](./docs/wd14_tagger_README-en.md) for details.
 - Fixed an error when specifying `--beam_search` and a value of 2 or more for `--num_beams` in `make_captions.py`.
-- The options `--noise_offset_random_strength` and `--ip_noise_gamma_random_strength` are added to each training script. These options can be used to vary the noise offset and ip noise gamma in the range of 0 to the specified value. PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) Thanks to KohakuBlueleaf!
-- The options `--save_state_on_train_end` are added to each training script. PR [#1168](https://github.com/kohya-ss/sd-scripts/pull/1168) Thanks to gesen2egee!
-- The options `--sample_every_n_epochs` and `--sample_every_n_steps` in each training script now display a warning and ignore them when a number less than or equal to `0` is specified. Thanks to S-Del for raising the issue.
-- The [English version of the dataset settings documentation](./docs/config_README-en.md) is added. PR [#1175](https://github.com/kohya-ss/sd-scripts/pull/1175) Thanks to darkstorm2150!
+
+#### About Masked loss
+
+The masked loss is supported in each training script. To enable the masked loss, specify the `--masked_loss` option.
+
+The feature is not fully tested, so there may be bugs. If you find any issues, please open an Issue.
+
+ControlNet dataset is used to specify the mask. The mask images should be the RGB images. The pixel value 255 in R channel is treated as the mask (the loss is calculated only for the pixels with the mask), and 0 is treated as the non-mask. The pixel values 0-255 are converted to 0-1 (i.e., the pixel value 128 is treated as the half weight of the loss). See details for the dataset specification in the [LLLite documentation](./docs/train_lllite_README.md#preparing-the-dataset).
+
+
+#### 主要な変更点
 
 - 依存ライブラリが更新されました。[アップグレード](./README-ja.md#アップグレード) を参照しライブラリを更新してください。
   - 特に `imagesize` が新しく追加されていますので、すぐにライブラリの更新ができない場合は `pip install imagesize==1.4.1` で個別にインストールしてください。
   - `bitsandbytes==0.43.0`、`prodigyopt==1.0`、`lion-pytorch==0.0.6` が requirements.txt に含まれるようになりました。
   - また PyTorch のバージョンを 2.1.2 に更新しました。PyTorch はすぐに更新する必要はありません。更新時は、アップグレードの手順では PyTorch が更新されませんので、torch、torchvision、xformers を手動でインストールしてください。
+- wandb へのログ出力が有効の場合、コマンドライン全体が公開されます。そのため、コマンドラインに wandb の API キーや HuggingFace のトークンなどが含まれる場合、設定ファイル（`.toml`）への記載をお勧めします。問題提起していただいた bghira 氏に感謝します。
+  - このような場合には学習開始時に警告が表示されます。
+  - また絶対パスの指定がある場合、そのパスが公開される可能性がありますので、相対パスを指定するか設定ファイルに記載することをお勧めします。このような場合は INFO ログが表示されます。
+  - 詳細は [#1123](https://github.com/kohya-ss/sd-scripts/pull/1123) および PR [#1240](https://github.com/kohya-ss/sd-scripts/pull/1240) をご覧ください。
 - Colab での動作時、ログ出力で停止してしまうようです。学習スクリプトに `--console_log_simple` オプションを指定し、rich のロギングを無効してお試しください。
-- データセット設定の `.toml` ファイルが UTF-8 encoding で読み込まれるようになりました。PR [#1167](https://github.com/kohya-ss/sd-scripts/pull/1167) Horizon1704 氏に感謝します。
-- データセット設定で、正則化画像のサブセットを複数指定した時、最後のサブセットの各種設定がすべてのサブセットの画像に適用される不具合が修正されました。それぞれのサブセットの設定が、それぞれの画像に正しく適用されます。PR [#1205](https://github.com/kohya-ss/sd-scripts/pull/1205) feffy380 氏に感謝します。
+- その他、マスクロス追加、DeepSpeed 対応、データセット設定の改善、画像タグ付けの改善などがあります。詳細は以下をご覧ください。
+
+#### 学習スクリプト
+
 - `train_network.py` および `sdxl_train_network.py` で、学習したモデルのメタデータに一部のデータセット設定が記録されるよう修正しました（`caption_prefix`、`caption_suffix`、`keep_tokens_separator`、`secondary_separator`、`enable_wildcard`）。
 - `train_network.py` および `sdxl_train_network.py` で、state に U-Net および Text Encoder が含まれる不具合を修正しました。state の保存、読み込みが高速化され、ファイルサイズも小さくなり、また読み込み時のメモリ使用量も削減されます。
 - DeepSpeed がサポートされました。PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101) 、[#1139](https://github.com/kohya-ss/sd-scripts/pull/1139) BootsofLagrangian 氏に感謝します。詳細は PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101) をご覧ください。
 - 各学習スクリプトでマスクロスをサポートしました。PR [#1207](https://github.com/kohya-ss/sd-scripts/pull/1207) 詳細は [Masked loss](#masked-loss) をご覧ください。
+- 各学習スクリプトに、noise offset、ip noise gammaを、それぞれ 0~指定した値の範囲で変動させるオプション `--noise_offset_random_strength` および `--ip_noise_gamma_random_strength` が追加されました。 PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) KohakuBlueleaf 氏に感謝します。
+- 各学習スクリプトに、学習終了時に state を保存する `--save_state_on_train_end` オプションが追加されました。 PR [#1168](https://github.com/kohya-ss/sd-scripts/pull/1168) gesen2egee 氏に感謝します。
+- 各学習スクリプトで `--sample_every_n_epochs` および `--sample_every_n_steps` オプションに `0` 以下の数値を指定した時、警告を表示するとともにそれらを無視するよう変更しました。問題提起していただいた S-Del 氏に感謝します。
+
+#### データセット設定
+
+- データセット設定の `.toml` ファイルが UTF-8 encoding で読み込まれるようになりました。PR [#1167](https://github.com/kohya-ss/sd-scripts/pull/1167) Horizon1704 氏に感謝します。
+- データセット設定で、正則化画像のサブセットを複数指定した時、最後のサブセットの各種設定がすべてのサブセットの画像に適用される不具合が修正されました。それぞれのサブセットの設定が、それぞれの画像に正しく適用されます。PR [#1205](https://github.com/kohya-ss/sd-scripts/pull/1205) feffy380 氏に感謝します。
 - データセットのサブセット設定にいくつかの機能を追加しました。
   - シャッフルの対象とならないタグ分割識別子の指定 `secondary_separator` を追加しました。`secondary_separator=";;;"` のように指定します。`secondary_separator` で区切ることで、その部分はシャッフル、drop 時にまとめて扱われます。
   - `enable_wildcard` を追加しました。`true` にするとワイルドカード記法 `{aaa|bbb|ccc}` が使えます。また複数行キャプションも有効になります。
@@ -194,6 +234,10 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
   - 既存の機能 `caption_prefix` と `caption_suffix` とあわせて使えます。`caption_prefix` と `caption_suffix` は一番最初に処理され、その後、ワイルドカード、`keep_tokens_separator`、シャッフルおよび drop、`secondary_separator` の順に処理されます。
   - 詳細は [データセット設定](./docs/config_README-ja.md) をご覧ください。
 - DreamBooth 方式の DataSet で画像情報（サイズ、キャプション）をキャッシュする機能が追加されました。PR [#1178](https://github.com/kohya-ss/sd-scripts/pull/1178)、[#1206](https://github.com/kohya-ss/sd-scripts/pull/1206) KohakuBlueleaf 氏に感謝します。詳細は [データセット設定](./docs/config_README-ja.md#dreambooth-方式専用のオプション) をご覧ください。
+- データセット設定の[英語版ドキュメント](./docs/config_README-en.md) が追加されました。PR [#1175](https://github.com/kohya-ss/sd-scripts/pull/1175) darkstorm2150 氏に感謝します。
+
+#### 画像のタグ付け
+
 - `tag_image_by_wd14_tagger.py` で v3 のリポジトリがサポートされました（`--onnx` 指定時のみ有効）。 PR [#1192](https://github.com/kohya-ss/sd-scripts/pull/1192) sdbds 氏に感謝します。
   - Onnx のバージョンアップが必要になるかもしれません。デフォルトでは Onnx はインストールされていませんので、`pip install onnx==1.15.0 onnxruntime-gpu==1.17.1` 等でインストール、アップデートしてください。`requirements.txt` のコメントもあわせてご確認ください。
 - `tag_image_by_wd14_tagger.py` で、モデルを`--repo_id` のサブディレクトリに保存するようにしました。これにより複数のモデルファイルがキャッシュされます。`--model_dir` 直下の不要なファイルは削除願います。
@@ -206,21 +250,8 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
   - タグを置換する `--tag_replacement`
   - 詳細は [タグ付けに関するドキュメント](./docs/wd14_tagger_README-ja.md) をご覧ください。
 - `make_captions.py` で `--beam_search` を指定し `--num_beams` に2以上の値を指定した時のエラーを修正しました。
-- 各学習スクリプトに、noise offset、ip noise gammaを、それぞれ 0~指定した値の範囲で変動させるオプション `--noise_offset_random_strength` および `--ip_noise_gamma_random_strength` が追加されました。 PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) KohakuBlueleaf 氏に感謝します。
-- 各学習スクリプトに、学習終了時に state を保存する `--save_state_on_train_end` オプションが追加されました。 PR [#1168](https://github.com/kohya-ss/sd-scripts/pull/1168) gesen2egee 氏に感謝します。
-- 各学習スクリプトで `--sample_every_n_epochs` および `--sample_every_n_steps` オプションに `0` 以下の数値を指定した時、警告を表示するとともにそれらを無視するよう変更しました。問題提起していただいた S-Del 氏に感謝します。
-- データセット設定の[英語版ドキュメント](./docs/config_README-en.md) が追加されました。PR [#1175](https://github.com/kohya-ss/sd-scripts/pull/1175) darkstorm2150 氏に感謝します。
-
-Please read [Releases](https://github.com/kohya-ss/sd-scripts/releases) for recent updates.
-最近の更新情報は [Release](https://github.com/kohya-ss/sd-scripts/releases) をご覧ください。
-
-#### Masked loss
 
-The masked loss is supported in each training script. To enable the masked loss, specify the `--masked_loss` option.
-
-The feature is not fully tested, so there may be bugs. If you find any issues, please open an Issue.
-
-ControlNet dataset is used to specify the mask. The mask images should be the RGB images. The pixel value 255 in R channel is treated as the mask (the loss is calculated only for the pixels with the mask), and 0 is treated as the non-mask. The pixel values 0-255 are converted to 0-1 (i.e., the pixel value 128 is treated as the half weight of the loss). See details for the dataset specification in the [LLLite documentation](./docs/train_lllite_README.md#preparing-the-dataset).
+#### マスクロスについて
 
 各学習スクリプトでマスクロスをサポートしました。マスクロスを有効にするには `--masked_loss` オプションを指定してください。
 
@@ -228,6 +259,8 @@ ControlNet dataset is used to specify the mask. The mask images should be the RG
 
 マスクの指定には ControlNet データセットを使用します。マスク画像は RGB 画像である必要があります。R チャンネルのピクセル値 255 がロス計算対象、0 がロス計算対象外になります。0-255 の値は、0-1 の範囲に変換されます（つまりピクセル値 128 の部分はロスの重みが半分になります）。データセットの詳細は [LLLite ドキュメント](./docs/train_lllite_README-ja.md#データセットの準備) をご覧ください。
 
+Please read [Releases](https://github.com/kohya-ss/sd-scripts/releases) for recent updates.
+最近の更新情報は [Release](https://github.com/kohya-ss/sd-scripts/releases) をご覧ください。
 
 ## Additional Information
 

From 90b18795fce516cb00735dc43a6ee76ecae8ec83 Mon Sep 17 00:00:00 2001
From: kabachuha <artemkhrapov2001@yandex.ru>
Date: Sun, 7 Apr 2024 07:54:21 +0300
Subject: [PATCH 67/69] Add option to use Scheduled Huber Loss in all training
 pipelines to improve resilience to data corruption (#1228)

* add huber loss and huber_c compute to train_util

* add reduction modes

* add huber_c retrieval from timestep getter

* move get timesteps and huber to own function

* add conditional loss to all training scripts

* add cond loss to train network

* add (scheduled) huber_loss to args

* fixup twice timesteps getting

* PHL-schedule should depend on noise scheduler's num timesteps

* *2 multiplier to huber loss cause of 1/2 a^2 conv.

The Taylor expansion of sqrt near zero gives 1/2 a^2, which differs from a^2 of the standard MSE loss. This change scales them better against one another

* add option for smooth l1 (huber / delta)

* unify huber scheduling

* add snr huber scheduler

---------

Co-authored-by: Kohya S <52813779+kohya-ss@users.noreply.github.com>
---
 fine_tune.py                         |  6 +--
 library/train_util.py                | 79 ++++++++++++++++++++++++++--
 sdxl_train.py                        |  6 +--
 sdxl_train_control_net_lllite.py     |  4 +-
 sdxl_train_control_net_lllite_old.py |  4 +-
 train_controlnet.py                  | 11 ++--
 train_db.py                          |  4 +-
 train_network.py                     |  4 +-
 train_textual_inversion.py           |  4 +-
 train_textual_inversion_XTI.py       |  4 +-
 10 files changed, 96 insertions(+), 30 deletions(-)

diff --git a/fine_tune.py b/fine_tune.py
index 3c4a5a26b..c7e6bbd2e 100644
--- a/fine_tune.py
+++ b/fine_tune.py
@@ -354,7 +354,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
 
                 # Sample noise, sample a random timestep for each image, and add noise to the latents,
                 # with noise offset and/or multires noise if specified
-                noise, noisy_latents, timesteps = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents)
+                noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents)
 
                 # Predict the noise residual
                 with accelerator.autocast():
@@ -368,7 +368,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
 
                 if args.min_snr_gamma or args.scale_v_pred_loss_like_noise_pred or args.debiased_estimation_loss:
                     # do not mean over batch dimension for snr weight or scale v-pred loss
-                    loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none")
+                    loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c)
                     loss = loss.mean([1, 2, 3])
 
                     if args.min_snr_gamma:
@@ -380,7 +380,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
 
                     loss = loss.mean()  # mean over batch dimension
                 else:
-                    loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="mean")
+                    loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="mean", loss_type=args.loss_type, huber_c=huber_c)
 
                 accelerator.backward(loss)
                 if accelerator.sync_gradients and args.max_grad_norm != 0.0:
diff --git a/library/train_util.py b/library/train_util.py
index c13bb68ee..90e6818ad 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -3236,6 +3236,26 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
         default=None,
         help="set maximum time step for U-Net training (1~1000, default is 1000) / U-Net学習時のtime stepの最大値を設定する（1~1000で指定、省略時はデフォルト値(1000)）",
     )
+    parser.add_argument(
+        "--loss_type",
+        type=str,
+        default="l2",
+        choices=["l2", "huber", "smooth_l1"],
+        help="The type of loss to use and whether it's scheduled based on the timestep"
+    )
+    parser.add_argument(
+        "--huber_schedule",
+        type=str,
+        default="exponential",
+        choices=["constant", "exponential", "snr"],
+        help="The type of loss to use and whether it's scheduled based on the timestep"
+    )
+    parser.add_argument(
+        "--huber_c",
+        type=float,
+        default=0.1,
+        help="The huber loss parameter. Only used if one of the huber loss modes (huber or smooth l1) is selected with loss_type.",
+    )
 
     parser.add_argument(
         "--lowram",
@@ -4842,6 +4862,38 @@ def save_sd_model_on_train_end_common(
         if args.huggingface_repo_id is not None:
             huggingface_util.upload(args, out_dir, "/" + model_name, force_sync_upload=True)
 
+def get_timesteps_and_huber_c(args, min_timestep, max_timestep, noise_scheduler, b_size, device):
+
+    #TODO: if a huber loss is selected, it will use constant timesteps for each batch
+    # as. In the future there may be a smarter way
+
+    if args.loss_type == 'huber' or args.loss_type == 'smooth_l1':
+        timesteps = torch.randint(
+            min_timestep, max_timestep, (1,), device='cpu'
+        )
+        timestep = timesteps.item()
+
+        if args.huber_schedule == "exponential":
+            alpha = - math.log(args.huber_c) / noise_scheduler.config.num_train_timesteps
+            huber_c = math.exp(-alpha * timestep)
+        elif args.huber_schedule == "snr":
+            alphas_cumprod = noise_scheduler.alphas_cumprod[timestep]
+            sigmas = ((1.0 - alphas_cumprod) / alphas_cumprod) ** 0.5
+            huber_c = (1 - args.huber_c) / (1 + sigmas)**2 + args.huber_c
+        elif args.huber_schedule == "constant":
+            huber_c = args.huber_c
+        else:
+            raise NotImplementedError(f'Unknown Huber loss schedule {args.huber_schedule}!')
+
+        timesteps = timesteps.repeat(b_size).to(device)
+    elif args.loss_type == 'l2':
+        timesteps = torch.randint(min_timestep, max_timestep, (b_size,), device=device)
+        huber_c = 1 # may be anything, as it's not used
+    else:
+        raise NotImplementedError(f'Unknown loss type {args.loss_type}')
+    timesteps = timesteps.long()
+
+    return timesteps, huber_c
 
 def get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents):
     # Sample noise that we'll add to the latents
@@ -4862,8 +4914,7 @@ def get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents):
     min_timestep = 0 if args.min_timestep is None else args.min_timestep
     max_timestep = noise_scheduler.config.num_train_timesteps if args.max_timestep is None else args.max_timestep
 
-    timesteps = torch.randint(min_timestep, max_timestep, (b_size,), device=latents.device)
-    timesteps = timesteps.long()
+    timesteps, huber_c = get_timesteps_and_huber_c(args, min_timestep, max_timestep, noise_scheduler, b_size, latents.device)
 
     # Add noise to the latents according to the noise magnitude at each timestep
     # (this is the forward diffusion process)
@@ -4876,8 +4927,28 @@ def get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents):
     else:
         noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
 
-    return noise, noisy_latents, timesteps
-
+    return noise, noisy_latents, timesteps, huber_c
+
+# NOTE: if you're using the scheduled version, huber_c has to depend on the timesteps already
+def conditional_loss(model_pred:torch.Tensor, target:torch.Tensor, reduction:str="mean", loss_type:str="l2", huber_c:float=0.1):
+    
+    if loss_type == 'l2':
+        loss = torch.nn.functional.mse_loss(model_pred, target, reduction=reduction)
+    elif loss_type == 'huber':
+        loss = 2 * huber_c * (torch.sqrt((model_pred - target) ** 2 + huber_c**2) - huber_c)
+        if reduction == "mean":
+            loss = torch.mean(loss)
+        elif reduction == "sum":
+            loss = torch.sum(loss)
+    elif loss_type == 'smooth_l1':
+        loss = 2 * (torch.sqrt((model_pred - target) ** 2 + huber_c**2) - huber_c)
+        if reduction == "mean":
+            loss = torch.mean(loss)
+        elif reduction == "sum":
+            loss = torch.sum(loss)
+    else:
+        raise NotImplementedError(f'Unsupported Loss Type {loss_type}')
+    return loss
 
 def append_lr_to_logs(logs, lr_scheduler, optimizer_type, including_unet=True):
     names = []
diff --git a/sdxl_train.py b/sdxl_train.py
index f6d277494..46d7860be 100644
--- a/sdxl_train.py
+++ b/sdxl_train.py
@@ -582,7 +582,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
 
                 # Sample noise, sample a random timestep for each image, and add noise to the latents,
                 # with noise offset and/or multires noise if specified
-                noise, noisy_latents, timesteps = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents)
+                noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents)
 
                 noisy_latents = noisy_latents.to(weight_dtype)  # TODO check why noisy_latents is not weight_dtype
 
@@ -600,7 +600,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
                     or args.masked_loss
                 ):
                     # do not mean over batch dimension for snr weight or scale v-pred loss
-                    loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none")
+                    loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c)
                     if args.masked_loss:
                         loss = apply_masked_loss(loss, batch)
                     loss = loss.mean([1, 2, 3])
@@ -616,7 +616,7 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
 
                     loss = loss.mean()  # mean over batch dimension
                 else:
-                    loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="mean")
+                    loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="mean", loss_type=args.loss_type, huber_c=huber_c)
 
                 accelerator.backward(loss)
                 if accelerator.sync_gradients and args.max_grad_norm != 0.0:
diff --git a/sdxl_train_control_net_lllite.py b/sdxl_train_control_net_lllite.py
index e880b57de..f89c3628f 100644
--- a/sdxl_train_control_net_lllite.py
+++ b/sdxl_train_control_net_lllite.py
@@ -439,7 +439,7 @@ def remove_model(old_ckpt_name):
 
                 # Sample noise, sample a random timestep for each image, and add noise to the latents,
                 # with noise offset and/or multires noise if specified
-                noise, noisy_latents, timesteps = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents)
+                noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents)
 
                 noisy_latents = noisy_latents.to(weight_dtype)  # TODO check why noisy_latents is not weight_dtype
 
@@ -458,7 +458,7 @@ def remove_model(old_ckpt_name):
                 else:
                     target = noise
 
-                loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none")
+                loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c)
                 loss = loss.mean([1, 2, 3])
 
                 loss_weights = batch["loss_weights"]  # 各sampleごとのweight
diff --git a/sdxl_train_control_net_lllite_old.py b/sdxl_train_control_net_lllite_old.py
index 0ea64b824..e85e978c1 100644
--- a/sdxl_train_control_net_lllite_old.py
+++ b/sdxl_train_control_net_lllite_old.py
@@ -406,7 +406,7 @@ def remove_model(old_ckpt_name):
 
                 # Sample noise, sample a random timestep for each image, and add noise to the latents,
                 # with noise offset and/or multires noise if specified
-                noise, noisy_latents, timesteps = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents)
+                noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents)
 
                 noisy_latents = noisy_latents.to(weight_dtype)  # TODO check why noisy_latents is not weight_dtype
 
@@ -426,7 +426,7 @@ def remove_model(old_ckpt_name):
                 else:
                     target = noise
 
-                loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none")
+                loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c)
                 loss = loss.mean([1, 2, 3])
 
                 loss_weights = batch["loss_weights"]  # 各sampleごとのweight
diff --git a/train_controlnet.py b/train_controlnet.py
index 90cac0410..f4c94e8d9 100644
--- a/train_controlnet.py
+++ b/train_controlnet.py
@@ -420,13 +420,8 @@ def remove_model(old_ckpt_name):
                     )
 
                 # Sample a random timestep for each image
-                timesteps = torch.randint(
-                    0,
-                    noise_scheduler.config.num_train_timesteps,
-                    (b_size,),
-                    device=latents.device,
-                )
-                timesteps = timesteps.long()
+                timesteps, huber_c = train_util.get_timesteps_and_huber_c(args, 0, noise_scheduler.config.num_train_timesteps, noise_scheduler, b_size, latents.device)
+
                 # Add noise to the latents according to the noise magnitude at each timestep
                 # (this is the forward diffusion process)
                 noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
@@ -457,7 +452,7 @@ def remove_model(old_ckpt_name):
                 else:
                     target = noise
 
-                loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none")
+                loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c)
                 loss = loss.mean([1, 2, 3])
 
                 loss_weights = batch["loss_weights"]  # 各sampleごとのweight
diff --git a/train_db.py b/train_db.py
index c3b7339f3..1de504ed8 100644
--- a/train_db.py
+++ b/train_db.py
@@ -346,7 +346,7 @@ def train(args):
 
                 # Sample noise, sample a random timestep for each image, and add noise to the latents,
                 # with noise offset and/or multires noise if specified
-                noise, noisy_latents, timesteps = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents)
+                noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents)
 
                 # Predict the noise residual
                 with accelerator.autocast():
@@ -358,7 +358,7 @@ def train(args):
                 else:
                     target = noise
 
-                loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none")
+                loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c)
                 if args.masked_loss:
                     loss = apply_masked_loss(loss, batch)
                 loss = loss.mean([1, 2, 3])
diff --git a/train_network.py b/train_network.py
index fcf4cd9b6..31d89276c 100644
--- a/train_network.py
+++ b/train_network.py
@@ -843,7 +843,7 @@ def remove_model(old_ckpt_name):
 
                     # Sample noise, sample a random timestep for each image, and add noise to the latents,
                     # with noise offset and/or multires noise if specified
-                    noise, noisy_latents, timesteps = train_util.get_noise_noisy_latents_and_timesteps(
+                    noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(
                         args, noise_scheduler, latents
                     )
 
@@ -873,7 +873,7 @@ def remove_model(old_ckpt_name):
                     else:
                         target = noise
 
-                    loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none")
+                    loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c)
                     if args.masked_loss:
                         loss = apply_masked_loss(loss, batch)
                     loss = loss.mean([1, 2, 3])
diff --git a/train_textual_inversion.py b/train_textual_inversion.py
index 02edf9525..10fce2677 100644
--- a/train_textual_inversion.py
+++ b/train_textual_inversion.py
@@ -572,7 +572,7 @@ def remove_model(old_ckpt_name):
 
                     # Sample noise, sample a random timestep for each image, and add noise to the latents,
                     # with noise offset and/or multires noise if specified
-                    noise, noisy_latents, timesteps = train_util.get_noise_noisy_latents_and_timesteps(
+                    noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(
                         args, noise_scheduler, latents
                     )
 
@@ -588,7 +588,7 @@ def remove_model(old_ckpt_name):
                     else:
                         target = noise
 
-                    loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none")
+                    loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c)
                     if args.masked_loss:
                         loss = apply_masked_loss(loss, batch)
                     loss = loss.mean([1, 2, 3])
diff --git a/train_textual_inversion_XTI.py b/train_textual_inversion_XTI.py
index f0723f2a7..ddd03d532 100644
--- a/train_textual_inversion_XTI.py
+++ b/train_textual_inversion_XTI.py
@@ -461,7 +461,7 @@ def remove_model(old_ckpt_name):
 
                 # Sample noise, sample a random timestep for each image, and add noise to the latents,
                 # with noise offset and/or multires noise if specified
-                noise, noisy_latents, timesteps = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents)
+                noise, noisy_latents, timesteps, huber_c = train_util.get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents)
 
                 # Predict the noise residual
                 with accelerator.autocast():
@@ -473,7 +473,7 @@ def remove_model(old_ckpt_name):
                 else:
                     target = noise
 
-                loss = torch.nn.functional.mse_loss(noise_pred.float(), target.float(), reduction="none")
+                loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c)
                 if args.masked_loss:
                     loss = apply_masked_loss(loss, batch)
                 loss = loss.mean([1, 2, 3])

From d30ebb205cbec29010d35222e2f478cf1813e151 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sun, 7 Apr 2024 14:58:17 +0900
Subject: [PATCH 68/69] update readme, add metadata for network module

---
 README.md             | 47 +++++++++++++++++++++++++++++++++++++++----
 library/train_util.py | 45 +++++++++++++++++++++++------------------
 train_network.py      | 11 ++++++++--
 3 files changed, 77 insertions(+), 26 deletions(-)

diff --git a/README.md b/README.md
index 0cecc5676..5282c1f69 100644
--- a/README.md
+++ b/README.md
@@ -150,14 +150,15 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
   - Also, if there is an absolute path, the path may be exposed, so it is recommended to specify a relative path or write it in the configuration file. In such cases, an INFO log is displayed.
   - See [#1123](https://github.com/kohya-ss/sd-scripts/pull/1123) and PR [#1240](https://github.com/kohya-ss/sd-scripts/pull/1240) for details.
 - Colab seems to stop with log output. Try specifying `--console_log_simple` option in the training script to disable rich logging.
-- Other improvements include the addition of masked loss, DeepSpeed support, dataset settings improvements, and image tagging improvements. See below for details.
+- Other improvements include the addition of masked loss, scheduled Huber Loss, DeepSpeed support, dataset settings improvements, and image tagging improvements. See below for details.
 
 #### Training scripts
 
 - `train_network.py` and `sdxl_train_network.py` are modified to record some dataset settings in the metadata of the trained model (`caption_prefix`, `caption_suffix`, `keep_tokens_separator`, `secondary_separator`, `enable_wildcard`).
 - Fixed a bug that U-Net and Text Encoders are included in the state in `train_network.py` and `sdxl_train_network.py`. The saving and loading of the state are faster, the file size is smaller, and the memory usage when loading is reduced.
 - DeepSpeed is supported. PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101)  and [#1139](https://github.com/kohya-ss/sd-scripts/pull/1139) Thanks to BootsofLagrangian! See PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101) for details.
-- The masked loss is supported in each training script. PR [#1207](https://github.com/kohya-ss/sd-scripts/pull/1207) See [Masked loss](#masked-loss) for details.
+- The masked loss is supported in each training script. PR [#1207](https://github.com/kohya-ss/sd-scripts/pull/1207) See [Masked loss](#about-masked-loss) for details.
+- Scheduled Huber Loss has been introduced to each training scripts. PR [#1228](https://github.com/kohya-ss/sd-scripts/pull/1228/) Thanks to kabachuha for the PR and cheald, drhead, and others for the discussion! See [Scheduled Huber Loss](#about-scheduled-huber-loss) for details.
 - The options `--noise_offset_random_strength` and `--ip_noise_gamma_random_strength` are added to each training script. These options can be used to vary the noise offset and ip noise gamma in the range of 0 to the specified value. PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) Thanks to KohakuBlueleaf!
 - The options `--save_state_on_train_end` are added to each training script. PR [#1168](https://github.com/kohya-ss/sd-scripts/pull/1168) Thanks to gesen2egee!
 - The options `--sample_every_n_epochs` and `--sample_every_n_steps` in each training script now display a warning and ignore them when a number less than or equal to `0` is specified. Thanks to S-Del for raising the issue.
@@ -199,6 +200,23 @@ The feature is not fully tested, so there may be bugs. If you find any issues, p
 
 ControlNet dataset is used to specify the mask. The mask images should be the RGB images. The pixel value 255 in R channel is treated as the mask (the loss is calculated only for the pixels with the mask), and 0 is treated as the non-mask. The pixel values 0-255 are converted to 0-1 (i.e., the pixel value 128 is treated as the half weight of the loss). See details for the dataset specification in the [LLLite documentation](./docs/train_lllite_README.md#preparing-the-dataset).
 
+#### About Scheduled Huber Loss
+
+Scheduled Huber Loss has been introduced to each training scripts. This is a method to improve robustness against outliers or anomalies (data corruption) in the training data.
+
+With the traditional MSE (L2) loss function, the impact of outliers could be significant, potentially leading to a degradation in the quality of generated images. On the other hand, while the Huber loss function can suppress the influence of outliers, it tends to compromise the reproduction of fine details in images.
+
+To address this, the proposed method employs a clever application of the Huber loss function. By scheduling the use of Huber loss in the early stages of training (when noise is high) and MSE in the later stages, it strikes a balance between outlier robustness and fine detail reproduction.
+
+Experimental results have confirmed that this method achieves higher accuracy on data containing outliers compared to pure Huber loss or MSE. The increase in computational cost is minimal.
+
+The newly added arguments loss_type, huber_schedule, and huber_c allow for the selection of the loss function type (Huber, smooth L1, MSE), scheduling method (exponential, constant, SNR), and Huber's parameter. This enables optimization based on the characteristics of the dataset.
+
+See PR [#1228](https://github.com/kohya-ss/sd-scripts/pull/1228/) for details.
+
+- `loss_type`: Specify the loss function type. Choose `huber` for Huber loss, `smooth_l1` for smooth L1 loss, and `l2` for MSE loss. The default is `l2`, which is the same as before.
+- `huber_schedule`: Specify the scheduling method. Choose `exponential`, `constant`, or `SNR`. The default is `exponential`.
+- `huber_c`: Specify the Huber's parameter. The default is `0.1`.
 
 #### 主要な変更点
 
@@ -211,14 +229,15 @@ ControlNet dataset is used to specify the mask. The mask images should be the RG
   - また絶対パスの指定がある場合、そのパスが公開される可能性がありますので、相対パスを指定するか設定ファイルに記載することをお勧めします。このような場合は INFO ログが表示されます。
   - 詳細は [#1123](https://github.com/kohya-ss/sd-scripts/pull/1123) および PR [#1240](https://github.com/kohya-ss/sd-scripts/pull/1240) をご覧ください。
 - Colab での動作時、ログ出力で停止してしまうようです。学習スクリプトに `--console_log_simple` オプションを指定し、rich のロギングを無効してお試しください。
-- その他、マスクロス追加、DeepSpeed 対応、データセット設定の改善、画像タグ付けの改善などがあります。詳細は以下をご覧ください。
+- その他、マスクロス追加、Scheduled Huber Loss 追加、DeepSpeed 対応、データセット設定の改善、画像タグ付けの改善などがあります。詳細は以下をご覧ください。
 
 #### 学習スクリプト
 
 - `train_network.py` および `sdxl_train_network.py` で、学習したモデルのメタデータに一部のデータセット設定が記録されるよう修正しました（`caption_prefix`、`caption_suffix`、`keep_tokens_separator`、`secondary_separator`、`enable_wildcard`）。
 - `train_network.py` および `sdxl_train_network.py` で、state に U-Net および Text Encoder が含まれる不具合を修正しました。state の保存、読み込みが高速化され、ファイルサイズも小さくなり、また読み込み時のメモリ使用量も削減されます。
 - DeepSpeed がサポートされました。PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101) 、[#1139](https://github.com/kohya-ss/sd-scripts/pull/1139) BootsofLagrangian 氏に感謝します。詳細は PR [#1101](https://github.com/kohya-ss/sd-scripts/pull/1101) をご覧ください。
-- 各学習スクリプトでマスクロスをサポートしました。PR [#1207](https://github.com/kohya-ss/sd-scripts/pull/1207) 詳細は [Masked loss](#masked-loss) をご覧ください。
+- 各学習スクリプトでマスクロスをサポートしました。PR [#1207](https://github.com/kohya-ss/sd-scripts/pull/1207) 詳細は [マスクロスについて](#マスクロスについて) をご覧ください。
+- 各学習スクリプトに Scheduled Huber Loss を追加しました。PR [#1228](https://github.com/kohya-ss/sd-scripts/pull/1228/) ご提案いただいた kabachuha 氏、および議論を深めてくださった cheald 氏、drhead 氏を始めとする諸氏に感謝します。詳細は [Scheduled Huber Loss について](#scheduled-huber-loss-について) をご覧ください。
 - 各学習スクリプトに、noise offset、ip noise gammaを、それぞれ 0~指定した値の範囲で変動させるオプション `--noise_offset_random_strength` および `--ip_noise_gamma_random_strength` が追加されました。 PR [#1177](https://github.com/kohya-ss/sd-scripts/pull/1177) KohakuBlueleaf 氏に感謝します。
 - 各学習スクリプトに、学習終了時に state を保存する `--save_state_on_train_end` オプションが追加されました。 PR [#1168](https://github.com/kohya-ss/sd-scripts/pull/1168) gesen2egee 氏に感謝します。
 - 各学習スクリプトで `--sample_every_n_epochs` および `--sample_every_n_steps` オプションに `0` 以下の数値を指定した時、警告を表示するとともにそれらを無視するよう変更しました。問題提起していただいた S-Del 氏に感謝します。
@@ -262,6 +281,26 @@ ControlNet dataset is used to specify the mask. The mask images should be the RG
 Please read [Releases](https://github.com/kohya-ss/sd-scripts/releases) for recent updates.
 最近の更新情報は [Release](https://github.com/kohya-ss/sd-scripts/releases) をご覧ください。
 
+#### Scheduled Huber Loss について
+
+各学習スクリプトに、学習データ中の異常値や外れ値（data corruption）への耐性を高めるための手法、Scheduled Huber Lossが導入されました。
+
+従来のMSE（L2）損失関数では、異常値の影響を大きく受けてしまい、生成画像の品質低下を招く恐れがありました。一方、Huber損失関数は異常値の影響を抑えられますが、画像の細部再現性が損なわれがちでした。
+
+この手法ではHuber損失関数の適用を工夫し、学習の初期段階（ノイズが大きい場合）ではHuber損失を、後期段階ではMSEを用いるようスケジューリングすることで、異常値耐性と細部再現性のバランスを取ります。
+
+実験の結果では、この手法が純粋なHuber損失やMSEと比べ、異常値を含むデータでより高い精度を達成することが確認されています。また計算コストの増加はわずかです。
+
+具体的には、新たに追加された引数loss_type、huber_schedule、huber_cで、損失関数の種類（Huber, smooth L1, MSE）とスケジューリング方法（exponential, constant, SNR）を選択できます。これによりデータセットに応じた最適化が可能になります。
+
+詳細は PR [#1228](https://github.com/kohya-ss/sd-scripts/pull/1228/) をご覧ください。
+
+- `loss_type` : 損失関数の種類を指定します。`huber` で Huber損失、`smooth_l1` で smooth L1 損失、`l2` で MSE 損失を選択します。デフォルトは `l2` で、従来と同様です。
+- `huber_schedule` : スケジューリング方法を指定します。`exponential` で指数関数的、`constant` で一定、`snr` で信号対雑音比に基づくスケジューリングを選択します。デフォルトは `exponential` です。
+- `huber_c` : Huber損失のパラメータを指定します。デフォルトは `0.1` です。
+
+PR 内でいくつかの比較が共有されています。この機能を試す場合、最初は `--loss_type smooth_l1 --huber_schedule snr --huber_c 0.1` などで試してみるとよいかもしれません。
+
 ## Additional Information
 
 ### Naming of LoRA
diff --git a/library/train_util.py b/library/train_util.py
index 90e6818ad..9ce129bd9 100644
--- a/library/train_util.py
+++ b/library/train_util.py
@@ -3241,20 +3241,21 @@ def add_training_arguments(parser: argparse.ArgumentParser, support_dreambooth:
         type=str,
         default="l2",
         choices=["l2", "huber", "smooth_l1"],
-        help="The type of loss to use and whether it's scheduled based on the timestep"
+        help="The type of loss function to use (L2, Huber, or smooth L1), default is L2 / 使用する損失関数の種類（L2、Huber、またはsmooth L1）、デフォルトはL2",
     )
     parser.add_argument(
         "--huber_schedule",
         type=str,
         default="exponential",
         choices=["constant", "exponential", "snr"],
-        help="The type of loss to use and whether it's scheduled based on the timestep"
+        help="The scheduling method for Huber loss (constant, exponential, or SNR-based). Only used when loss_type is 'huber' or 'smooth_l1'. default is exponential"
+        + " / Huber損失のスケジューリング方法（constant、exponential、またはSNRベース）。loss_typeが'huber'または'smooth_l1'の場合に有効、デフォルトはexponential",
     )
     parser.add_argument(
         "--huber_c",
         type=float,
         default=0.1,
-        help="The huber loss parameter. Only used if one of the huber loss modes (huber or smooth l1) is selected with loss_type.",
+        help="The huber loss parameter. Only used if one of the huber loss modes (huber or smooth l1) is selected with loss_type. default is 0.1 / Huber損失のパラメータ。loss_typeがhuberまたはsmooth l1の場合に有効。デフォルトは0.1",
     )
 
     parser.add_argument(
@@ -4862,39 +4863,39 @@ def save_sd_model_on_train_end_common(
         if args.huggingface_repo_id is not None:
             huggingface_util.upload(args, out_dir, "/" + model_name, force_sync_upload=True)
 
+
 def get_timesteps_and_huber_c(args, min_timestep, max_timestep, noise_scheduler, b_size, device):
 
-    #TODO: if a huber loss is selected, it will use constant timesteps for each batch
+    # TODO: if a huber loss is selected, it will use constant timesteps for each batch
     # as. In the future there may be a smarter way
 
-    if args.loss_type == 'huber' or args.loss_type == 'smooth_l1':
-        timesteps = torch.randint(
-            min_timestep, max_timestep, (1,), device='cpu'
-        )
+    if args.loss_type == "huber" or args.loss_type == "smooth_l1":
+        timesteps = torch.randint(min_timestep, max_timestep, (1,), device="cpu")
         timestep = timesteps.item()
 
         if args.huber_schedule == "exponential":
-            alpha = - math.log(args.huber_c) / noise_scheduler.config.num_train_timesteps
+            alpha = -math.log(args.huber_c) / noise_scheduler.config.num_train_timesteps
             huber_c = math.exp(-alpha * timestep)
         elif args.huber_schedule == "snr":
             alphas_cumprod = noise_scheduler.alphas_cumprod[timestep]
             sigmas = ((1.0 - alphas_cumprod) / alphas_cumprod) ** 0.5
-            huber_c = (1 - args.huber_c) / (1 + sigmas)**2 + args.huber_c
+            huber_c = (1 - args.huber_c) / (1 + sigmas) ** 2 + args.huber_c
         elif args.huber_schedule == "constant":
             huber_c = args.huber_c
         else:
-            raise NotImplementedError(f'Unknown Huber loss schedule {args.huber_schedule}!')
+            raise NotImplementedError(f"Unknown Huber loss schedule {args.huber_schedule}!")
 
         timesteps = timesteps.repeat(b_size).to(device)
-    elif args.loss_type == 'l2':
+    elif args.loss_type == "l2":
         timesteps = torch.randint(min_timestep, max_timestep, (b_size,), device=device)
-        huber_c = 1 # may be anything, as it's not used
+        huber_c = 1  # may be anything, as it's not used
     else:
-        raise NotImplementedError(f'Unknown loss type {args.loss_type}')
+        raise NotImplementedError(f"Unknown loss type {args.loss_type}")
     timesteps = timesteps.long()
 
     return timesteps, huber_c
 
+
 def get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents):
     # Sample noise that we'll add to the latents
     noise = torch.randn_like(latents, device=latents.device)
@@ -4929,27 +4930,31 @@ def get_noise_noisy_latents_and_timesteps(args, noise_scheduler, latents):
 
     return noise, noisy_latents, timesteps, huber_c
 
+
 # NOTE: if you're using the scheduled version, huber_c has to depend on the timesteps already
-def conditional_loss(model_pred:torch.Tensor, target:torch.Tensor, reduction:str="mean", loss_type:str="l2", huber_c:float=0.1):
-    
-    if loss_type == 'l2':
+def conditional_loss(
+    model_pred: torch.Tensor, target: torch.Tensor, reduction: str = "mean", loss_type: str = "l2", huber_c: float = 0.1
+):
+
+    if loss_type == "l2":
         loss = torch.nn.functional.mse_loss(model_pred, target, reduction=reduction)
-    elif loss_type == 'huber':
+    elif loss_type == "huber":
         loss = 2 * huber_c * (torch.sqrt((model_pred - target) ** 2 + huber_c**2) - huber_c)
         if reduction == "mean":
             loss = torch.mean(loss)
         elif reduction == "sum":
             loss = torch.sum(loss)
-    elif loss_type == 'smooth_l1':
+    elif loss_type == "smooth_l1":
         loss = 2 * (torch.sqrt((model_pred - target) ** 2 + huber_c**2) - huber_c)
         if reduction == "mean":
             loss = torch.mean(loss)
         elif reduction == "sum":
             loss = torch.sum(loss)
     else:
-        raise NotImplementedError(f'Unsupported Loss Type {loss_type}')
+        raise NotImplementedError(f"Unsupported Loss Type {loss_type}")
     return loss
 
+
 def append_lr_to_logs(logs, lr_scheduler, optimizer_type, including_unet=True):
     names = []
     if including_unet:
diff --git a/train_network.py b/train_network.py
index 31d89276c..c99d37247 100644
--- a/train_network.py
+++ b/train_network.py
@@ -476,7 +476,7 @@ def save_model_hook(models, weights, output_dir):
             # pop weights of other models than network to save only network weights
             if accelerator.is_main_process:
                 remove_indices = []
-                for i,model in enumerate(models):
+                for i, model in enumerate(models):
                     if not isinstance(model, type(accelerator.unwrap_model(network))):
                         remove_indices.append(i)
                 for i in reversed(remove_indices):
@@ -569,6 +569,11 @@ def load_model_hook(models, input_dir):
             "ss_scale_weight_norms": args.scale_weight_norms,
             "ss_ip_noise_gamma": args.ip_noise_gamma,
             "ss_debiased_estimation": bool(args.debiased_estimation_loss),
+            "ss_noise_offset_random_strength": args.noise_offset_random_strength,
+            "ss_ip_noise_gamma_random_strength": args.ip_noise_gamma_random_strength,
+            "ss_loss_type": args.loss_type,
+            "ss_huber_schedule": args.huber_schedule,
+            "ss_huber_c": args.huber_c,
         }
 
         if use_user_config:
@@ -873,7 +878,9 @@ def remove_model(old_ckpt_name):
                     else:
                         target = noise
 
-                    loss = train_util.conditional_loss(noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c)
+                    loss = train_util.conditional_loss(
+                        noise_pred.float(), target.float(), reduction="none", loss_type=args.loss_type, huber_c=huber_c
+                    )
                     if args.masked_loss:
                         loss = apply_masked_loss(loss, batch)
                     loss = loss.mean([1, 2, 3])

From dfa30790a99754f2f20e4980e91a000cce51d8a4 Mon Sep 17 00:00:00 2001
From: Kohya S <ykumeykume@gmail.com>
Date: Sun, 7 Apr 2024 20:34:26 +0900
Subject: [PATCH 69/69] update readme

---
 README-ja.md | 44 ++++++++++++++++++++++++++++++++++++++++++
 README.md    | 54 ++++++++++------------------------------------------
 2 files changed, 54 insertions(+), 44 deletions(-)

diff --git a/README-ja.md b/README-ja.md
index f70f882d7..4ae6b2334 100644
--- a/README-ja.md
+++ b/README-ja.md
@@ -111,3 +111,47 @@ Conv2d 3x3への拡大は [cloneofsimo氏](https://github.com/cloneofsimo/lora)
 
 [BLIP](https://github.com/salesforce/BLIP): BSD-3-Clause
 
+## その他の情報
+
+### LoRAの名称について
+
+`train_network.py` がサポートするLoRAについて、混乱を避けるため名前を付けました。ドキュメントは更新済みです。以下は当リポジトリ内の独自の名称です。
+
+1. __LoRA-LierLa__ : (LoRA for __Li__ n __e__ a __r__  __La__ yers、リエラと読みます)
+
+    Linear 層およびカーネルサイズ 1x1 の Conv2d 層に適用されるLoRA
+
+2. __LoRA-C3Lier__ : (LoRA for __C__ olutional layers with __3__ x3 Kernel and  __Li__ n __e__ a __r__ layers、セリアと読みます)
+
+    1.に加え、カーネルサイズ 3x3 の Conv2d 層に適用されるLoRA
+
+デフォルトではLoRA-LierLaが使われます。LoRA-C3Lierを使う場合は `--network_args` に `conv_dim` を指定してください。
+
+<!-- 
+LoRA-LierLa は[Web UI向け拡張](https://github.com/kohya-ss/sd-webui-additional-networks)、またはAUTOMATIC1111氏のWeb UIのLoRA機能で使用することができます。
+
+LoRA-C3Lierを使いWeb UIで生成するには拡張を使用してください。
+-->
+
+### 学習中のサンプル画像生成
+
+プロンプトファイルは例えば以下のようになります。
+
+```
+# prompt 1
+masterpiece, best quality, (1girl), in white shirts, upper body, looking at viewer, simple background --n low quality, worst quality, bad anatomy,bad composition, poor, low effort --w 768 --h 768 --d 1 --l 7.5 --s 28
+
+# prompt 2
+masterpiece, best quality, 1boy, in business suit, standing at street, looking back --n (low quality, worst quality), bad anatomy,bad composition, poor, low effort --w 576 --h 832 --d 2 --l 5.5 --s 40
+```
+
+  `#` で始まる行はコメントになります。`--n` のように「ハイフン二個＋英小文字」の形でオプションを指定できます。以下が使用可能できます。
+
+  * `--n` Negative prompt up to the next option.
+  * `--w` Specifies the width of the generated image.
+  * `--h` Specifies the height of the generated image.
+  * `--d` Specifies the seed of the generated image.
+  * `--l` Specifies the CFG scale of the generated image.
+  * `--s` Specifies the number of steps in the generation.
+
+  `( )` や `[ ]` などの重みづけも動作します。
diff --git a/README.md b/README.md
index 5282c1f69..1ca699be5 100644
--- a/README.md
+++ b/README.md
@@ -137,15 +137,16 @@ The majority of scripts is licensed under ASL 2.0 (including codes from Diffuser
 
 ## Change History
 
-### Mar XX, 2024 / 2024/3/XX: v0.8.6
+### Apr 7, 2024 / 2024-04-07: v0.8.6
 
 #### Highlights
 
 - The dependent libraries are updated. Please see [Upgrade](#upgrade) and update the libraries.
   - Especially `imagesize` is newly added, so if you cannot update the libraries immediately, please install with `pip install imagesize==1.4.1` separately.
   - `bitsandbytes==0.43.0`, `prodigyopt==1.0`, `lion-pytorch==0.0.6` are included in the requirements.txt.
+    - `bitsandbytes` no longer requires complex procedures as it now officially supports Windows.  
   - Also, the PyTorch version is updated to 2.1.2 (PyTorch does not need to be updated immediately). In the upgrade procedure, PyTorch is not updated, so please manually install or update torch, torchvision, xformers if necessary (see [Upgrade PyTorch](#upgrade-pytorch)).
-- When logging to wandb is enabled, the entire command line is exposed. Therefore, it is recommended to write the API key of wandb and the token of HuggingFace in the configuration file (`.toml`). Thanks to bghira for raising the issue.
+- When logging to wandb is enabled, the entire command line is exposed. Therefore, it is recommended to write wandb API key and HuggingFace token in the configuration file (`.toml`). Thanks to bghira for raising the issue.
   - A warning is displayed at the start of training if such information is included in the command line.
   - Also, if there is an absolute path, the path may be exposed, so it is recommended to specify a relative path or write it in the configuration file. In such cases, an INFO log is displayed.
   - See [#1123](https://github.com/kohya-ss/sd-scripts/pull/1123) and PR [#1240](https://github.com/kohya-ss/sd-scripts/pull/1240) for details.
@@ -223,6 +224,7 @@ See PR [#1228](https://github.com/kohya-ss/sd-scripts/pull/1228/) for details.
 - 依存ライブラリが更新されました。[アップグレード](./README-ja.md#アップグレード) を参照しライブラリを更新してください。
   - 特に `imagesize` が新しく追加されていますので、すぐにライブラリの更新ができない場合は `pip install imagesize==1.4.1` で個別にインストールしてください。
   - `bitsandbytes==0.43.0`、`prodigyopt==1.0`、`lion-pytorch==0.0.6` が requirements.txt に含まれるようになりました。
+    - `bitsandbytes` が公式に Windows をサポートしたため複雑な手順が不要になりました。
   - また PyTorch のバージョンを 2.1.2 に更新しました。PyTorch はすぐに更新する必要はありません。更新時は、アップグレードの手順では PyTorch が更新されませんので、torch、torchvision、xformers を手動でインストールしてください。
 - wandb へのログ出力が有効の場合、コマンドライン全体が公開されます。そのため、コマンドラインに wandb の API キーや HuggingFace のトークンなどが含まれる場合、設定ファイル（`.toml`）への記載をお勧めします。問題提起していただいた bghira 氏に感謝します。
   - このような場合には学習開始時に警告が表示されます。
@@ -315,27 +317,14 @@ The LoRA supported by `train_network.py` has been named to avoid confusion. The
 
     In addition to 1., LoRA for Conv2d layers with 3x3 kernel 
     
-LoRA-LierLa is the default LoRA type for `train_network.py` (without `conv_dim` network arg). LoRA-LierLa can be used with [our extension](https://github.com/kohya-ss/sd-webui-additional-networks) for AUTOMATIC1111's Web UI, or with the built-in LoRA feature of the Web UI.
-
-To use LoRA-C3Lier with Web UI, please use our extension.
-
-### LoRAの名称について
-
-`train_network.py` がサポートするLoRAについて、混乱を避けるため名前を付けました。ドキュメントは更新済みです。以下は当リポジトリ内の独自の名称です。
-
-1. __LoRA-LierLa__ : (LoRA for __Li__ n __e__ a __r__  __La__ yers、リエラと読みます)
-
-    Linear 層およびカーネルサイズ 1x1 の Conv2d 層に適用されるLoRA
-
-2. __LoRA-C3Lier__ : (LoRA for __C__ olutional layers with __3__ x3 Kernel and  __Li__ n __e__ a __r__ layers、セリアと読みます)
-
-    1.に加え、カーネルサイズ 3x3 の Conv2d 層に適用されるLoRA
-
-LoRA-LierLa は[Web UI向け拡張](https://github.com/kohya-ss/sd-webui-additional-networks)、またはAUTOMATIC1111氏のWeb UIのLoRA機能で使用することができます。
+LoRA-LierLa is the default LoRA type for `train_network.py` (without `conv_dim` network arg). 
+<!-- 
+LoRA-LierLa can be used with [our extension](https://github.com/kohya-ss/sd-webui-additional-networks) for AUTOMATIC1111's Web UI, or with the built-in LoRA feature of the Web UI.
 
-LoRA-C3Lierを使いWeb UIで生成するには拡張を使用してください。
+To use LoRA-C3Lier with Web UI, please use our extension. 
+-->
 
-## Sample image generation during training
+### Sample image generation during training
   A prompt file might look like this, for example
 
 ```
@@ -356,26 +345,3 @@ masterpiece, best quality, 1boy, in business suit, standing at street, looking b
   * `--s` Specifies the number of steps in the generation.
 
   The prompt weighting such as `( )` and `[ ]` are working.
-
-## サンプル画像生成
-プロンプトファイルは例えば以下のようになります。
-
-```
-# prompt 1
-masterpiece, best quality, (1girl), in white shirts, upper body, looking at viewer, simple background --n low quality, worst quality, bad anatomy,bad composition, poor, low effort --w 768 --h 768 --d 1 --l 7.5 --s 28
-
-# prompt 2
-masterpiece, best quality, 1boy, in business suit, standing at street, looking back --n (low quality, worst quality), bad anatomy,bad composition, poor, low effort --w 576 --h 832 --d 2 --l 5.5 --s 40
-```
-
-  `#` で始まる行はコメントになります。`--n` のように「ハイフン二個＋英小文字」の形でオプションを指定できます。以下が使用可能できます。
-
-  * `--n` Negative prompt up to the next option.
-  * `--w` Specifies the width of the generated image.
-  * `--h` Specifies the height of the generated image.
-  * `--d` Specifies the seed of the generated image.
-  * `--l` Specifies the CFG scale of the generated image.
-  * `--s` Specifies the number of steps in the generation.
-
-  `( )` や `[ ]` などの重みづけも動作します。
-