v0.1.2

README.md

@@ -11,15 +11,15 @@
 
 ## Install
 1. Create a virtual environment and install Python (>= 3.8)
-2. Download the latest version (>=1.12) of Torch(corresponding CUDA version) from the [official website](https://pytorch.org/get-started/locally/) of Torch. It is not recommended to automatically install Torch (CUDA 10.2) using the Mini-Lightning dependency, which will cause CUDA version mismatch.
+2. Download the latest version (>=1.12) of Torch(corresponding CUDA version) from the [official website](https://pytorch.org/get-started/locally/) of Torch. It is not recommended to automatically install Torch (CUDA 10.2 default) using the Mini-Lightning dependency, which will cause CUDA version mismatch.
 3. Install mini-lightning
 ```bash
-# from pypi (v0.1.1)
+# from pypi (v0.1.2)
 pip install mini-lightning
 
 # Or download the files from the repository to local,
 # and go to the folder where setup.py is located, and run the following command
-# (recommend) You can enjoy the latest features and functions
+# (Recommended) You can enjoy the latest features and functions (including bug fixes)
 pip install .
 ```
 
@@ -47,7 +47,7 @@ pip install gym, pygame
 python examples/dqn.py
 
 ### cv_ddp.py; cv_ddp_spawn.py
-# torchrun (recommended): Ref: https://pytorch.org/docs/stable/elastic/run.html
+# torchrun (Recommended): Ref: https://pytorch.org/docs/stable/elastic/run.html
 # spawn: Ref: https://pytorch.org/docs/stable/notes/ddp.html
 ## single-gpu  # for test
 torchrun examples/cv_ddp.py --device_ids 0

examples/cv.py

@@ -8,11 +8,10 @@
 import torchvision.models as tvm
 #
 CIFAR10 = tvd.CIFAR10
-RUNS_DIR = os.path.join(RUNS_DIR, "cv")
 DATASETS_PATH = os.environ.get("DATASETS_PATH", os.path.join(RUNS_DIR, "datasets"))
-CHECKPOINTS_PATH = os.path.join(RUNS_DIR, "checkpoints")
+RUNS_DIR = os.path.join(RUNS_DIR, "cv")
 os.makedirs(DATASETS_PATH, exist_ok=True)
-os.makedirs(CHECKPOINTS_PATH, exist_ok=True)
+os.makedirs(RUNS_DIR, exist_ok=True)
 
 #
 device_ids = [0]
@@ -116,7 +115,6 @@ def test_step(self, batch: Any) -> None:
     ldm = ml.LDataModule(
         train_dataset, val_dataset, test_dataset, **hparams["dataloader_hparams"])
 
-    runs_dir = CHECKPOINTS_PATH
     loss_fn = nn.CrossEntropyLoss()
 
     def collect_res(seed: int) -> Dict[str, float]:
@@ -129,7 +127,7 @@ def collect_res(seed: int) -> Dict[str, float]:
         lr_s = ml.WarmupCosineAnnealingLR(optimizer, **hparams["lrs_hparams"])
 
         lmodel = MyLModule(model, optimizer, loss_fn, lr_s, hparams)
-        trainer = ml.Trainer(lmodel, device_ids, runs_dir=runs_dir, **hparams["trainer_hparams"])
+        trainer = ml.Trainer(lmodel, device_ids, runs_dir=RUNS_DIR, **hparams["trainer_hparams"])
         res = trainer.fit(ldm.train_dataloader, ldm.val_dataloader)
         res2 = trainer.test(ldm.test_dataloader)
         res.update(res2)

examples/cv_ddp.py

@@ -13,11 +13,10 @@
 import torchvision.models as tvm
 #
 CIFAR10 = tvd.CIFAR10
-RUNS_DIR = os.path.join(RUNS_DIR, "cv_ddp")
 DATASETS_PATH = os.environ.get("DATASETS_PATH", os.path.join(RUNS_DIR, "datasets"))
-CHECKPOINTS_PATH = os.path.join(RUNS_DIR, "checkpoints")
+RUNS_DIR = os.path.join(RUNS_DIR, "cv_ddp")
 os.makedirs(DATASETS_PATH, exist_ok=True)
-os.makedirs(CHECKPOINTS_PATH, exist_ok=True)
+os.makedirs(RUNS_DIR, exist_ok=True)
 #
 
 
@@ -130,7 +129,6 @@ def parse_opt() -> Namespace:
     ldm = ml.LDataModule(
         train_dataset, val_dataset, test_dataset, **hparams["dataloader_hparams"])
 
-    runs_dir = CHECKPOINTS_PATH
     loss_fn = nn.CrossEntropyLoss()
 
     def collect_res(seed: int) -> Dict[str, float]:
@@ -146,7 +144,7 @@ def collect_res(seed: int) -> Dict[str, float]:
         lr_s = ml.WarmupCosineAnnealingLR(optimizer, **hparams["lrs_hparams"])
         #
         lmodel = MyLModule(model, optimizer, loss_fn, lr_s, hparams)
-        trainer = ml.Trainer(lmodel, device_ids, runs_dir=runs_dir, **hparams["trainer_hparams"])
+        trainer = ml.Trainer(lmodel, device_ids, runs_dir=RUNS_DIR, **hparams["trainer_hparams"])
         res = trainer.fit(ldm.train_dataloader, ldm.val_dataloader)
         res2 = trainer.test(ldm.test_dataloader)
         res.update(res2)

examples/cv_ddp_spawn.py

@@ -13,11 +13,10 @@
 import torchvision.models as tvm
 #
 CIFAR10 = tvd.CIFAR10
-RUNS_DIR = os.path.join(RUNS_DIR, "cv_ddp")
 DATASETS_PATH = os.environ.get("DATASETS_PATH", os.path.join(RUNS_DIR, "datasets"))
-CHECKPOINTS_PATH = os.path.join(RUNS_DIR, "checkpoints")
+RUNS_DIR = os.path.join(RUNS_DIR, "cv_ddp")
 os.makedirs(DATASETS_PATH, exist_ok=True)
-os.makedirs(CHECKPOINTS_PATH, exist_ok=True)
+os.makedirs(RUNS_DIR, exist_ok=True)
 #
 
 
@@ -125,7 +124,6 @@ def main(rank: int, world_size: int, device_ids: List[int]) -> None:
     ldm = ml.LDataModule(
         train_dataset, val_dataset, test_dataset, **hparams["dataloader_hparams"])
 
-    runs_dir = CHECKPOINTS_PATH
     loss_fn = nn.CrossEntropyLoss()
 
     def collect_res(seed: int) -> Dict[str, float]:
@@ -141,7 +139,7 @@ def collect_res(seed: int) -> Dict[str, float]:
         lr_s = ml.WarmupCosineAnnealingLR(optimizer, **hparams["lrs_hparams"])
         #
         lmodel = MyLModule(model, optimizer, loss_fn, lr_s, hparams)
-        trainer = ml.Trainer(lmodel, device_ids, runs_dir=runs_dir, **hparams["trainer_hparams"])
+        trainer = ml.Trainer(lmodel, device_ids, runs_dir=RUNS_DIR, **hparams["trainer_hparams"])
         res = trainer.fit(ldm.train_dataloader, ldm.val_dataloader)
         res2 = trainer.test(ldm.test_dataloader)
         res.update(res2)

examples/dqn.py

@@ -18,11 +18,10 @@
 
 
 RENDER = True
-RUNS_DIR = os.path.join(RUNS_DIR, "dqn")
 DATASETS_PATH = os.environ.get("DATASETS_PATH", os.path.join(RUNS_DIR, "datasets"))
-CHECKPOINTS_PATH = os.path.join(RUNS_DIR, "checkpoints")
+RUNS_DIR = os.path.join(RUNS_DIR, "dqn")
 os.makedirs(DATASETS_PATH, exist_ok=True)
-os.makedirs(CHECKPOINTS_PATH, exist_ok=True)
+os.makedirs(RUNS_DIR, exist_ok=True)
 
 #
 device_ids = [0]
@@ -231,9 +230,8 @@ def training_step(self, batch: Any) -> Tensor:
     #
     get_rand_p = partial(get_rand_p, **hparams["rand_p"])
     optimizer = getattr(optim, hparams["optim_name"])(model.parameters(), **hparams["optim_hparams"])
-    runs_dir = CHECKPOINTS_PATH
     loss_fn = nn.MSELoss()
 
     lmodel = MyLModule(model, optimizer, loss_fn, agent, get_rand_p, hparams)
-    trainer = ml.Trainer(lmodel, device_ids, runs_dir=runs_dir, **hparams["trainer_hparams"])
+    trainer = ml.Trainer(lmodel, device_ids, runs_dir=RUNS_DIR, **hparams["trainer_hparams"])
     trainer.fit(ldm.train_dataloader, ldm.val_dataloader)

examples/nlp.py

@@ -11,11 +11,10 @@
 
 
 #
-RUNS_DIR = os.path.join(RUNS_DIR, "nlp")
 DATASETS_PATH = os.environ.get("DATASETS_PATH", os.path.join(RUNS_DIR, "datasets"))
-CHECKPOINTS_PATH = os.path.join(RUNS_DIR, "checkpoints")
+RUNS_DIR = os.path.join(RUNS_DIR, "nlp")
 os.makedirs(DATASETS_PATH, exist_ok=True)
-os.makedirs(CHECKPOINTS_PATH, exist_ok=True)
+os.makedirs(RUNS_DIR, exist_ok=True)
 os.environ["TOKENIZERS_PARALLELISM"] = "true"
 
 #
@@ -113,11 +112,10 @@ def tokenize_function(example):
         "recall": Recall(average="macro", num_classes=2),
         "f1": F1Score(average="none", num_classes=2)
     }
-    runs_dir = CHECKPOINTS_PATH
     loss_fn = nn.CrossEntropyLoss()
     lr_s = ml.WarmupCosineAnnealingLR(optimizer, **hparams["lrs_hparams"])
     lmodel = MyLModule(model, optimizer, metrics, loss_fn, lr_s, hparams)
-    trainer = ml.Trainer(lmodel, device_ids, runs_dir=runs_dir, **hparams["trainer_hparams"])
+    trainer = ml.Trainer(lmodel, device_ids, runs_dir=RUNS_DIR, **hparams["trainer_hparams"])
     try:
         logger.info(trainer.fit(ldm.train_dataloader, ldm.val_dataloader))
     except KeyboardInterrupt:

setup.py

@@ -21,7 +21,7 @@ def read_file(path: str) -> str:
 ]
 setup(
     name="mini-lightning",
-    version="0.1.2.dev",
+    version="0.1.2",
     description=description,
     long_description=long_description,
     long_description_content_type='text/markdown',