Merge pull request #865 from bmaltais/dev2

v21.5.14
bmaltais · May 28, 2023 · 47b55a4 · 47b55a4
2 parents 30b054b + 6bb8ec7
commit 47b55a4
Show file tree

Hide file tree

Showing 14 changed files with 443 additions and 86 deletions.
diff --git a/README.md b/README.md
@@ -345,15 +345,25 @@ This will store a backup file with your current locally installed pip packages a
 
 ## Change History
 
-* 2023/07/15 (v21.5.12)
+* 2023/05/28 (v21.5.14)
+- Add Create Groupo tool and GUI
+* 2023/05/24 (v21.5.13)
+- Upgrade gradio release to fix issue with UI refresh on config load.
+- [D-Adaptation v3.0](https://github.com/facebookresearch/dadaptation) is now supported. [PR #530](https://github.com/kohya-ss/sd-scripts/pull/530) Thanks to sdbds!
+  - `--optimizer_type` now accepts `DAdaptAdamPreprint`, `DAdaptAdanIP`, and `DAdaptLion`.
+  - `DAdaptAdam` is now new. The old `DAdaptAdam` is available with `DAdaptAdamPreprint`.
+  - Simply specifying `DAdaptation` will use `DAdaptAdamPreprint` (same behavior as before).
+  - You need to install D-Adaptation v3.0. After activating venv, please do `pip install -U dadaptation`.
+  - See PR and D-Adaptation documentation for details.
+* 2023/05/22 (v21.5.12)
 - Fixed several bugs.
   - The state is saved even when the `--save_state` option is not specified in `fine_tune.py` and `train_db.py`. [PR #521](https://github.com/kohya-ss/sd-scripts/pull/521) Thanks to akshaal!
   - Cannot load LoRA without `alpha`. [PR #527](https://github.com/kohya-ss/sd-scripts/pull/527) Thanks to Manjiz!
   - Minor changes to console output during sample generation. [PR #515](https://github.com/kohya-ss/sd-scripts/pull/515) Thanks to yanhuifair!
 - The generation script now uses xformers for VAE as well.
 - Fixed an issue where an error would occur if the encoding of the prompt file was different from the default. [PR #510](https://github.com/kohya-ss/sd-scripts/pull/510) Thanks to sdbds!
   - Please save the prompt file in UTF-8.
-* 2023/07/15 (v21.5.11)
+* 2023/05/15 (v21.5.11)
   - Added an option `--dim_from_weights` to `train_network.py` to automatically determine the dim(rank) from the weight file. [PR #491](https://github.com/kohya-ss/sd-scripts/pull/491) Thanks to AI-Casanova!
     - It is useful in combination with `resize_lora.py`. Please see the PR for details.
   - Fixed a bug where the noise resolution was incorrect with Multires noise. [PR #489](https://github.com/kohya-ss/sd-scripts/pull/489) Thanks to sdbds!

diff --git a/docs/train_README-ja.md b/docs/train_README-ja.md
@@ -615,9 +615,12 @@ masterpiece, best quality, 1boy, in business suit, standing at street, looking b
     - Lion8bit : 引数は同上
     - SGDNesterov : [torch.optim.SGD](https://pytorch.org/docs/stable/generated/torch.optim.SGD.html), nesterov=True
     - SGDNesterov8bit : 引数は同上
-    - DAdaptation(DAdaptAdam) : https://github.com/facebookresearch/dadaptation
+    - DAdaptation(DAdaptAdamPreprint) : https://github.com/facebookresearch/dadaptation
+    - DAdaptAdam : 引数は同上
     - DAdaptAdaGrad : 引数は同上
     - DAdaptAdan : 引数は同上
+    - DAdaptAdanIP : 引数は同上
+    - DAdaptLion : 引数は同上
     - DAdaptSGD : 引数は同上
     - AdaFactor : [Transformers AdaFactor](https://huggingface.co/docs/transformers/main_classes/optimizer_schedules)
     - 任意のオプティマイザ

diff --git a/docs/train_README-zh.md b/docs/train_README-zh.md
@@ -550,8 +550,14 @@ masterpiece, best quality, 1boy, in business suit, standing at street, looking b
     - Lion : https://github.com/lucidrains/lion-pytorch
     - 与过去版本中指定的 --use_lion_optimizer 相同
     - SGDNesterov : [torch.optim.SGD](https://pytorch.org/docs/stable/generated/torch.optim.SGD.html), nesterov=True
-    - SGDNesterov8bit : 引数同上
-    - DAdaptation : https://github.com/facebookresearch/dadaptation
+    - SGDNesterov8bit : 参数同上
+    - DAdaptation(DAdaptAdamPreprint) : https://github.com/facebookresearch/dadaptation
+    - DAdaptAdam : 参数同上
+    - DAdaptAdaGrad : 参数同上
+    - DAdaptAdan : 参数同上
+    - DAdaptAdanIP : 引数は同上
+    - DAdaptLion : 参数同上
+    - DAdaptSGD : 参数同上
     - AdaFactor : [Transformers AdaFactor](https://huggingface.co/docs/transformers/main_classes/optimizer_schedules)
     - 任何优化器
 

diff --git a/docs/train_network_README-ja.md b/docs/train_network_README-ja.md
@@ -276,7 +276,9 @@ python networks\merge_lora.py --sd_model ..\model\model.ckpt
 
 ### 複数のLoRAのモデルをマージする
 
-複数のLoRAモデルをひとつずつSDモデルに適用する場合と、複数のLoRAモデルをマージしてからSDモデルにマージする場合とは、計算順序の関連で微妙に異なる結果になります。
+__複数のLoRAをマージする場合は原則として `svd_merge_lora.py` を使用してください。__ 単純なup同士やdown同士のマージでは、計算結果が正しくなくなるためです。
+
+`merge_lora.py` によるマージは差分抽出法でLoRAを生成する場合等、ごく限られた場合でのみ有効です。
 
 たとえば以下のようなコマンドラインになります。
 
@@ -294,7 +296,7 @@ python networks\merge_lora.py
 
 --ratiosにそれぞれのモデルの比率（どのくらい重みを元モデルに反映するか）を0~1.0の数値で指定します。二つのモデルを一対一でマージす場合は、「0.5 0.5」になります。「1.0 1.0」では合計の重みが大きくなりすぎて、恐らく結果はあまり望ましくないものになると思われます。
 
-v1で学習したLoRAとv2で学習したLoRA、rank（次元数）や``alpha``の異なるLoRAはマージできません。U-NetだけのLoRAとU-Net+Text EncoderのLoRAはマージできるはずですが、結果は未知数です。
+v1で学習したLoRAとv2で学習したLoRA、rank（次元数）の異なるLoRAはマージできません。U-NetだけのLoRAとU-Net+Text EncoderのLoRAはマージできるはずですが、結果は未知数です。
 
 
 ### その他のオプション

diff --git a/kohya_gui.py b/kohya_gui.py
@@ -118,12 +118,13 @@ def UI(**kwargs):
                 enable_copy_info_button=True,
                 headless=headless,
             )
-            gradio_extract_dylora_tab(headless=headless)
-            gradio_extract_lora_tab(headless=headless)
-            gradio_extract_lycoris_locon_tab(headless=headless)
-            gradio_merge_lora_tab(headless=headless)
-            gradio_merge_lycoris_tab(headless=headless)
-            gradio_resize_lora_tab(headless=headless)
+            with gr.Tab('LoRA'):
+                gradio_extract_dylora_tab(headless=headless)
+                gradio_extract_lora_tab(headless=headless)
+                gradio_extract_lycoris_locon_tab(headless=headless)
+                gradio_merge_lora_tab(headless=headless)
+                gradio_merge_lycoris_tab(headless=headless)
+                gradio_resize_lora_tab(headless=headless)
 
     # Show the interface
     launch_kwargs = {}

diff --git a/library/common_gui.py b/library/common_gui.py
@@ -786,7 +786,11 @@ def gradio_training(
                 'Adafactor',
                 'DAdaptation',
                 'DAdaptAdaGrad',
+                'DAdaptAdam',
                 'DAdaptAdan',
+                'DAdaptAdanIP',
+                'DAdaptAdamPreprint',
+                'DAdaptLion',
                 'DAdaptSGD',
                 'Lion',
                 'Lion8bit',

diff --git a/library/group_images_gui.py b/library/group_images_gui.py
@@ -0,0 +1,110 @@
+import gradio as gr
+from easygui import msgbox
+import subprocess
+from .common_gui import get_folder_path
+import os
+
+PYTHON = 'python3' if os.name == 'posix' else './venv/Scripts/python.exe'
+
+def group_images(
+    input_folder,
+    output_folder,
+    group_size,
+    include_subfolders,
+    do_not_copy_other_files
+):
+    if input_folder == '':
+        msgbox('Input folder is missing...')
+        return
+
+    if output_folder == '':
+        msgbox('Please provide an output folder.')
+        return
+
+    print(f'Grouping images in {input_folder}...')
+
+    run_cmd = f'{PYTHON} "{os.path.join("tools","group_images.py")}"'
+    run_cmd += f' "{input_folder}"'
+    run_cmd += f' "{output_folder}"'
+    run_cmd += f' {(group_size)}'
+    if include_subfolders:
+        run_cmd += f' --include_subfolders'
+    if do_not_copy_other_files:
+        run_cmd += f' --do_not_copy_other_files'
+
+    print(run_cmd)
+
+    if os.name == 'posix':
+        os.system(run_cmd)
+    else:
+        subprocess.run(run_cmd)
+
+    print('...grouping done')
+
+
+def gradio_group_images_gui_tab(headless=False):
+    with gr.Tab('Group Images'):
+        gr.Markdown('This utility will group images in a folder based on their aspect ratio.')
+
+        with gr.Row():
+            input_folder = gr.Textbox(
+                label='Input folder',
+                placeholder='Directory containing the images to group',
+                interactive=True,
+            )
+            button_input_folder = gr.Button(
+                '📂', elem_id='open_folder_small', visible=(not headless)
+            )
+            button_input_folder.click(
+                get_folder_path,
+                outputs=input_folder,
+                show_progress=False,
+            )
+
+            output_folder = gr.Textbox(
+                label='Output folder',
+                placeholder='Directory where the grouped images will be stored',
+                interactive=True,
+            )
+            button_output_folder = gr.Button(
+                '📂', elem_id='open_folder_small', visible=(not headless)
+            )
+            button_output_folder.click(
+                get_folder_path,
+                outputs=output_folder,
+                show_progress=False,
+            )
+        with gr.Row():
+            group_size = gr.Slider(
+                label='Group size',
+                info='Number of images to group together',
+                value='4',
+                minimum=1, maximum=64, step=1,
+                interactive=True,
+            )
+
+            include_subfolders = gr.Checkbox(
+                label='Include Subfolders',
+                value=False,
+                info='Include images in subfolders as well',
+            )
+
+            do_not_copy_other_files = gr.Checkbox(
+                label='Do not copy other files',
+                value=False,
+                info='Do not copy other files in the input folder to the output folder',
+            )
+
+        group_images_button = gr.Button('Group images')
+
+        group_images_button.click(
+            group_images,
+            inputs=[
+                input_folder,
+                output_folder,
+                group_size,
+                include_subfolders,
+                do_not_copy_other_files
+            ],
+            show_progress=False,
+        )
diff --git a/library/train_util.py b/library/train_util.py
@@ -1940,7 +1940,7 @@ def add_optimizer_arguments(parser: argparse.ArgumentParser):
         "--optimizer_type",
         type=str,
         default="",
-        help="Optimizer to use / オプティマイザの種類: AdamW (default), AdamW8bit, Lion8bit, Lion, SGDNesterov, SGDNesterov8bit, DAdaptation(DAdaptAdam), DAdaptAdaGrad, DAdaptAdan, DAdaptSGD, AdaFactor",
+        help="Optimizer to use / オプティマイザの種類: AdamW (default), AdamW8bit, Lion8bit, Lion, SGDNesterov, SGDNesterov8bit, DAdaptation(DAdaptAdamPreprint), DAdaptAdaGrad, DAdaptAdam, DAdaptAdan, DAdaptAdanIP, DAdaptLion, DAdaptSGD, AdaFactor",
     )
 
     # backward compatibility
@@ -2545,7 +2545,7 @@ def task():
 
 
 def get_optimizer(args, trainable_params):
-    # "Optimizer to use: AdamW, AdamW8bit, Lion, SGDNesterov, SGDNesterov8bit, Lion8bit, DAdaptation, DAdaptation(DAdaptAdam), DAdaptAdaGrad, DAdaptAdan, DAdaptSGD, Adafactor"
+    # "Optimizer to use: AdamW, AdamW8bit, Lion, SGDNesterov, SGDNesterov8bit, Lion8bit, DAdaptation(DAdaptAdamPreprint), DAdaptAdaGrad, DAdaptAdam, DAdaptAdan, DAdaptAdanIP, DAdaptLion, DAdaptSGD, Adafactor"
 
     optimizer_type = args.optimizer_type
     if args.use_8bit_adam:
@@ -2653,6 +2653,7 @@ def get_optimizer(args, trainable_params):
         # check dadaptation is installed
         try:
             import dadaptation
+            import dadaptation.experimental as experimental
         except ImportError:
             raise ImportError("No dadaptation / dadaptation がインストールされていないようです")
 
@@ -2677,15 +2678,24 @@ def get_optimizer(args, trainable_params):
             )
 
         # set optimizer
-        if optimizer_type == "DAdaptation".lower() or optimizer_type == "DAdaptAdam".lower():
-            optimizer_class = dadaptation.DAdaptAdam
-            print(f"use D-Adaptation Adam optimizer | {optimizer_kwargs}")
+        if optimizer_type == "DAdaptation".lower() or optimizer_type == "DAdaptAdamPreprint".lower():
+            optimizer_class = experimental.DAdaptAdamPreprint
+            print(f"use D-Adaptation AdamPreprint optimizer | {optimizer_kwargs}")
         elif optimizer_type == "DAdaptAdaGrad".lower():
             optimizer_class = dadaptation.DAdaptAdaGrad
             print(f"use D-Adaptation AdaGrad optimizer | {optimizer_kwargs}")
+        elif optimizer_type == "DAdaptAdam".lower():
+            optimizer_class = dadaptation.DAdaptAdam
+            print(f"use D-Adaptation Adam optimizer | {optimizer_kwargs}")
         elif optimizer_type == "DAdaptAdan".lower():
             optimizer_class = dadaptation.DAdaptAdan
             print(f"use D-Adaptation Adan optimizer | {optimizer_kwargs}")
+        elif optimizer_type == "DAdaptAdanIP".lower():
+            optimizer_class = experimental.DAdaptAdanIP
+            print(f"use D-Adaptation AdanIP optimizer | {optimizer_kwargs}")
+        elif optimizer_type == "DAdaptLion".lower():
+            optimizer_class = dadaptation.DAdaptLion
+            print(f"use D-Adaptation Lion optimizer | {optimizer_kwargs}")
         elif optimizer_type == "DAdaptSGD".lower():
             optimizer_class = dadaptation.DAdaptSGD
             print(f"use D-Adaptation SGD optimizer | {optimizer_kwargs}")

diff --git a/library/utilities.py b/library/utilities.py
@@ -11,6 +11,7 @@
 from library.blip_caption_gui import gradio_blip_caption_gui_tab
 from library.git_caption_gui import gradio_git_caption_gui_tab
 from library.wd14_caption_gui import gradio_wd14_caption_gui_tab
+from library.group_images_gui import gradio_group_images_gui_tab
 
 
 def utilities_tab(
@@ -28,6 +29,7 @@ def utilities_tab(
         gradio_git_caption_gui_tab(headless=headless)
         gradio_wd14_caption_gui_tab(headless=headless)
     gradio_convert_model_tab(headless=headless)
+    gradio_group_images_gui_tab(headless=headless)
 
     return (
         train_data_dir_input,

diff --git a/lora_gui.py b/lora_gui.py
@@ -558,6 +558,12 @@ def train_model(
         )
         reg_factor = 2
 
+    print(f'Total steps: {total_steps}')
+    print(f'Train batch size: {train_batch_size}')
+    print(f'Gradient accumulation steps: {gradient_accumulation_steps}')
+    print(f'Epoch: {epoch}')
+    print(f'Regulatization factor: {reg_factor}')
+
     # calculate max_train_steps
     max_train_steps = int(
         math.ceil(
@@ -568,7 +574,7 @@ def train_model(
             * int(reg_factor)
         )
     )
-    print(f'max_train_steps = {max_train_steps}')
+    print(f'max_train_steps ({total_steps} / {train_batch_size} / {gradient_accumulation_steps} * {epoch} * {reg_factor}) = {max_train_steps}')
 
     # calculate stop encoder training
     if stop_text_encoder_training_pct == None:

diff --git a/requirements.txt b/requirements.txt
@@ -6,12 +6,12 @@ altair==4.2.2
 # This next line is not an error but rather there to properly catch if the url based bitsandbytes was properly installed by the line above...
 bitsandbytes==0.35.0; sys_platform == 'win32'
 bitsandbytes==0.38.1; (sys_platform == "darwin" or sys_platform == "linux")
-dadaptation==1.5
+dadaptation==3.1
 diffusers[torch]==0.10.2
 easygui==0.98.3
 einops==0.6.0
 ftfy==6.1.1
-gradio==3.28.1; sys_platform != 'darwin'
+gradio==3.32.0; sys_platform != 'darwin'
 gradio==3.23.0; sys_platform == 'darwin'
 lion-pytorch==0.0.6
 opencv-python==4.7.0.68

diff --git a/setup.bat b/setup.bat
@@ -42,7 +42,7 @@ if %choice%==1 (
     pip install --use-pep517 --upgrade -r requirements.txt
     pip install -U -I --no-deps https://github.com/C43H66N12O12S2/stable-diffusion-webui/releases/download/f/xformers-0.0.14.dev0-cp310-cp310-win_amd64.whl
 ) else (
-    pip install torch==2.0.0+cu118 torchvision==0.15.1+cu118 --extra-index-url https://download.pytorch.org/whl/cu118
+    pip install torch==2.0.1+cu118 torchvision==0.15.2+cu118 --extra-index-url https://download.pytorch.org/whl/cu118
     pip install --use-pep517 --upgrade -r requirements.txt
     pip install --upgrade xformers==0.0.19
     rem pip install -U -I --no-deps https://files.pythonhosted.org/packages/d6/f7/02662286419a2652c899e2b3d1913c47723fc164b4ac06a85f769c291013/xformers-0.0.17rc482-cp310-cp310-win_amd64.whl