commit

README.md

@@ -48,7 +48,7 @@ pip install "transformers>=4.23.*" "datasets>=2.6.*"
 python examples/nlp.py
 
 ### dqn.py
-pip install "gym>=0.26.*" pygame
+pip install "gym>=0.26.2" "pygame>=2.1.2"
 python examples/dqn.py
 
 ### gan.py

examples/audio.py

@@ -0,0 +1,5 @@
+# Author: Jintao Huang
+# Email: [email protected]
+# Date:
+
+# Ref: https://pytorch.org/audio/stable/index.html

examples/contrastive_learning.py

@@ -20,6 +20,7 @@ class Acc(MeanMetric):
     higher_is_better = True
     full_state_update = False
 
+
 #
 device_ids = [0]
 
@@ -58,7 +59,7 @@ def __init__(self, hparams: Dict[str, Any]) -> None:
         #
         resnet.fc = nn.Sequential(
             resnet.fc,
-            nn.ReLU(),
+            nn.ReLU(inplace=True),
             nn.Linear(4*hidden_state, hidden_state)
         )
         optimizer: Optimizer = getattr(optim, hparams["optim_name"])(resnet.parameters(), **hparams["optim_hparams"])
@@ -110,10 +111,10 @@ def test_step(self, batch: Any) -> None:
 
 class MLP(ml.LModule):
     def __init__(self, resnet: Module, hparams: Dict[str, Any]) -> None:
-        hidden_state = hparams["hidden_state"]
+        in_channels = hparams["in_channels"]
         n_classes = hparams["n_classes"]
         #
-        mlp = nn.Linear(4*hidden_state, n_classes)
+        mlp = nn.Linear(in_channels, n_classes)
         optimizer: Optimizer = getattr(optim, hparams["optim_name"])(mlp.parameters(), **hparams["optim_hparams"])
         lr_s: LRScheduler = lrs.CosineAnnealingLR(optimizer, **hparams["lrs_hparams"])
         metrics = {
@@ -173,7 +174,7 @@ def transforms(x: Image.Image) -> List[Tensor]:
             tvt.ToTensor(),
             tvt.Normalize((0.5,), (0.5,)),
         ])
-        return [contrast_transforms(x) for i in range(n_views)]
+        return [contrast_transforms(x) for _ in range(n_views)]
 
     train_dataset = STL10(
         root=DATASETS_PATH,
@@ -224,6 +225,7 @@ def transforms(x: Image.Image) -> List[Tensor]:
     trainer = ml.Trainer(lmodel, device_ids, runs_dir=RUNS_DIR, **hparams["trainer_hparams"])
     logger.info(trainer.fit(ldm.train_dataloader, ldm.val_dataloader))
     resnet = deepcopy(lmodel.resnet)
+    in_channels = resnet.fc[0].in_features
     resnet.fc = nn.Identity()
     ##########
     del ldm, lmodel, trainer, transforms, train_dataset, val_dataset, max_epochs, batch_size, hparams
@@ -232,7 +234,7 @@ def transforms(x: Image.Image) -> List[Tensor]:
     batch_size = 256
     hparams = {
         "device_ids": device_ids,
-        "hidden_state": hidden_state,
+        "in_channels": in_channels,
         "n_classes": 10,
         "dataloader_hparams": {"batch_size": batch_size, "num_workers": 4},
         "optim_name": "SGD",

examples/cv_ddp.py

@@ -114,7 +114,7 @@ def parse_opt() -> Namespace:
         "model_name": "resnet50",
         "model_hparams": {"num_classes": 10},
         "model_pretrain_model": {"url": tvm.ResNet50_Weights.DEFAULT.url},
-        "dataloader_hparams": {"batch_size": batch_size, "num_workers": 4, "pin_memory_train": False},
+        "dataloader_hparams": {"batch_size": batch_size, "num_workers": 4},  # "pin_memory": False
         "optim_name": "SGD",
         "optim_hparams": {"lr": 1e-2, "weight_decay": 1e-4, "momentum": 0.9},
         "trainer_hparams": {

mini_lightning/_mini_lightning.py

@@ -272,15 +272,19 @@ def __init__(
         #
         batch_size: int,
         num_workers: int = 0,
+        pin_memory: bool = True,
         collate_fn: Optional[Callable[[List[Any]], Any]] = None,  # for test/val and (train if collate_fn_train is None)
         *,
         shuffle_train: bool = True,
-        pin_memory_train: bool = True,
         drop_last_train: bool = True,  # If DP/DDP, drop_last=False may cause uneven split
+        pin_memory_train: Optional[bool] = None,
         collate_fn_train: Optional[Callable[[List[Any]], Any]] = None,
     ) -> None:
+        if pin_memory_train is None:
+            pin_memory_train = pin_memory
         if collate_fn_train is None:
             collate_fn_train = collate_fn
+        #
         self.train_dataloader: Optional[DataLoader] = None
         self.val_dataloader: Optional[DataLoader] = None
         self.test_dataloader: Optional[DataLoader] = None
@@ -294,7 +298,7 @@ def __init__(
         for dataset, loader_name in zip([val_dataset, test_dataset], ["val_dataloader", "test_dataloader"]):
             if rank in {-1, 0} and dataset is not None:
                 loader = DataLoader(dataset, batch_size, shuffle=False,
-                                    num_workers=num_workers, pin_memory=False,
+                                    num_workers=num_workers, pin_memory=pin_memory,
                                     drop_last=False, collate_fn=collate_fn)
                 setattr(self, loader_name, loader)