v0.1.7

README.md

@@ -56,7 +56,7 @@ pip install "torchvision>=0.13.*"
 python examples/gan.py
 
 ### contrastive_learning.py
-pip install "torchvision>=0.13.*" "torchmetrics>=0.10.2" "scikit-learn>=1.1.*"
+pip install "torchvision>=0.13.*" "scikit-learn>=1.1.*"
 python examples/contrastive_learning.py
 
 ### gnn.py gnn2.py

examples/test_env.py

@@ -140,7 +140,7 @@ def training_epoch_end(self) -> Dict[str, float]:
     ldm = ml.LDataModule(train_dataset, val_dataset, test_dataset, 64)
     lmodel = MyLModule(model, [optimizer], "loss")
     trainer = ml.Trainer(lmodel, [0], 20, RUNS_DIR, gradient_clip_norm=10,
-                         val_every_n_epoch=10, verbose=True, model_fpath=ckpt_path)
+                         val_every_n_epoch=10, verbose=True, ckpt_fpath=ckpt_path)
     logger.info(trainer.test(ldm.val_dataloader, True, True))
     logger.info(trainer.fit(ldm.train_dataloader, ldm.val_dataloader))
     logger.info(trainer.test(ldm.test_dataloader, True, True))
@@ -152,5 +152,5 @@ def training_epoch_end(self) -> Dict[str, float]:
     model = MLP_L2(2, 4, 1)
     ldm = ml.LDataModule(train_dataset, val_dataset, test_dataset, 64)
     lmodel = MyLModule(model, [], "loss")
-    trainer = ml.Trainer(lmodel, [], None, RUNS_DIR, model_fpath=ckpt_path)
+    trainer = ml.Trainer(lmodel, [], None, RUNS_DIR, ckpt_fpath=ckpt_path)
     logger.info(trainer.test(ldm.test_dataloader, True, True))

mini_lightning/_mini_lightning.py

@@ -313,7 +313,7 @@ def __init__(
         gradient_clip_norm: Optional[float] = None,
         sync_bn: bool = False,
         replace_sampler_ddp: bool = True,
-        model_fpath: Optional[str] = None,
+        ckpt_fpath: Optional[str] = None,
         #
         val_every_n_epoch: int = 1,
         log_every_n_steps: int = 10,
@@ -360,7 +360,7 @@ def __init__(
             replace_sampler_ddp=False: each gpu will use the complete dataset.
             replace_sampler_ddp=True: It will slice the dataset into world_size chunks and distribute them to each gpu.
             note: Replace train_dataloader only. Because DDP uses a single gpu for val/test. 
-        model_fpath: only load model_state_dict. 
+        ckpt_fpath: only load model_state_dict. 
             If you want to resume from ckpt. please see `save_optimizers_state_dict` and examples in `examples/test_env.py`
         *
         val_every_n_epoch: Frequency of validation and prog_bar_leave of training. (the last epoch will always be validated)
@@ -472,9 +472,10 @@ def __init__(
             hparams = lmodel.hparams
             self.save_hparams(hparams)
         #
-        if model_fpath is not None:
-            self._load_ckpt(model_fpath)
-            logger.info(f"Using ckpt: {model_fpath}")
+        self.ckpt_fpath = ckpt_fpath
+        if ckpt_fpath is not None:
+            self._load_ckpt(ckpt_fpath)
+            logger.info(f"Using ckpt: {ckpt_fpath}")
         lmodel.trainer_init(self)
         for s in lmodel._models:
             model: Module = getattr(lmodel, s)