Upgrade PyTorch Lightning to 1.0.2

examples/lightning_base.py

@@ -337,7 +337,7 @@ def add_generic_args(parser, root_dir) -> None:
 def generic_train(
     model: BaseTransformer,
     args: argparse.Namespace,
-    early_stopping_callback=False,
+    early_stopping_callback=None,
     logger=True,  # can pass WandbLogger() here
     extra_callbacks=[],
     checkpoint_callback=None,
@@ -355,6 +355,8 @@ def generic_train(
         checkpoint_callback = pl.callbacks.ModelCheckpoint(
             filepath=args.output_dir, prefix="checkpoint", monitor="val_loss", mode="min", save_top_k=1
         )
+    if early_stopping_callback:
+        extra_callbacks.append(early_stopping_callback)
     if logging_callback is None:
         logging_callback = LoggingCallback()
 
@@ -376,7 +378,6 @@ def generic_train(
         callbacks=[logging_callback] + extra_callbacks,
         logger=logger,
         checkpoint_callback=checkpoint_callback,
-        early_stop_callback=early_stopping_callback,
         **train_params,
     )
 

examples/requirements.txt

@@ -5,7 +5,7 @@ psutil
 sacrebleu
 rouge-score
 tensorflow_datasets
-pytorch-lightning==0.9.0
+pytorch-lightning==1.0.2
 matplotlib
 git-python==1.0.3
 faiss-cpu

examples/seq2seq/callbacks.py

@@ -102,7 +102,6 @@ def get_checkpoint_callback(output_dir, metric, save_top_k=1, lower_is_better=Fa
         monitor=f"val_{metric}",
         mode="min" if "loss" in metric else "max",
         save_top_k=save_top_k,
-        period=0,  # maybe save a checkpoint every time val is run, not just end of epoch.
     )
     return checkpoint_callback
 

examples/text-classification/run_pl_glue.py

@@ -192,7 +192,7 @@ def main():
 
     # Optionally, predict on dev set and write to output_dir
     if args.do_predict:
-        checkpoints = list(sorted(glob.glob(os.path.join(args.output_dir, "checkpointepoch=*.ckpt"), recursive=True)))
+        checkpoints = list(sorted(glob.glob(os.path.join(args.output_dir, "checkpoint-epoch=*.ckpt"), recursive=True)))
         model = model.load_from_checkpoint(checkpoints[-1])
         return trainer.test(model)
 

examples/token-classification/run_pl_ner.py

@@ -207,9 +207,9 @@ def add_model_specific_args(parser, root_dir):
 
     if args.do_predict:
         # See https://github.com/huggingface/transformers/issues/3159
-        # pl use this format to create a checkpoint:
+        # pl use this default format to create a checkpoint:
         # https://github.com/PyTorchLightning/pytorch-lightning/blob/master\
-        # /pytorch_lightning/callbacks/model_checkpoint.py#L169
-        checkpoints = list(sorted(glob.glob(os.path.join(args.output_dir, "checkpointepoch=*.ckpt"), recursive=True)))
+        # /pytorch_lightning/callbacks/model_checkpoint.py#L322
+        checkpoints = list(sorted(glob.glob(os.path.join(args.output_dir, "checkpoint-epoch=*.ckpt"), recursive=True)))
         model = model.load_from_checkpoint(checkpoints[-1])
         trainer.test(model)