Split checkpoint from model_name_or_path in examples (huggingface#11492)

* Split checkpoint from model_name_or_path in examples

* Address review comments

* Address review comments

examples/pytorch/README.md

@@ -65,7 +65,7 @@ examples/pytorch/token-classification/run_ner.py -h
 You can resume training from a previous checkpoint like this:
 
 1. Pass `--output_dir previous_output_dir` without `--overwrite_output_dir` to resume training from the latest checkpoint in `output_dir` (what you would use if the training was interrupted, for instance).
-2. Pass `--model_name_or_path path_to_a_specific_checkpoint` to resume training from that checkpoint folder.
+2. Pass `--resume_from_checkpoint path_to_a_specific_checkpoint` to resume training from that checkpoint folder.
 
 Should you want to turn an example into a notebook where you'd no longer have access to the command
 line, 🤗 Trainer supports resuming from a checkpoint via `trainer.train(resume_from_checkpoint)`.

examples/pytorch/language-modeling/run_clm.py

@@ -190,7 +190,7 @@ def main():
                 f"Output directory ({training_args.output_dir}) already exists and is not empty. "
                 "Use --overwrite_output_dir to overcome."
             )
-        elif last_checkpoint is not None:
+        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
             logger.info(
                 f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
                 "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
@@ -413,12 +413,11 @@ def group_texts(examples):
 
     # Training
     if training_args.do_train:
-        if last_checkpoint is not None:
+        checkpoint = None
+        if training_args.resume_from_checkpoint is not None:
+            checkpoint = training_args.resume_from_checkpoint
+        elif last_checkpoint is not None:
             checkpoint = last_checkpoint
-        elif model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path):
-            checkpoint = model_args.model_name_or_path
-        else:
-            checkpoint = None
         train_result = trainer.train(resume_from_checkpoint=checkpoint)
         trainer.save_model()  # Saves the tokenizer too for easy upload
 

examples/pytorch/language-modeling/run_mlm.py

@@ -199,7 +199,7 @@ def main():
                 f"Output directory ({training_args.output_dir}) already exists and is not empty. "
                 "Use --overwrite_output_dir to overcome."
             )
-        elif last_checkpoint is not None:
+        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
             logger.info(
                 f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
                 "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
@@ -443,12 +443,11 @@ def group_texts(examples):
 
     # Training
     if training_args.do_train:
-        if last_checkpoint is not None:
+        checkpoint = None
+        if training_args.resume_from_checkpoint is not None:
+            checkpoint = training_args.resume_from_checkpoint
+        elif last_checkpoint is not None:
             checkpoint = last_checkpoint
-        elif model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path):
-            checkpoint = model_args.model_name_or_path
-        else:
-            checkpoint = None
         train_result = trainer.train(resume_from_checkpoint=checkpoint)
         trainer.save_model()  # Saves the tokenizer too for easy upload
         metrics = train_result.metrics

examples/pytorch/language-modeling/run_plm.py

@@ -196,7 +196,7 @@ def main():
                 f"Output directory ({training_args.output_dir}) already exists and is not empty. "
                 "Use --overwrite_output_dir to overcome."
             )
-        elif last_checkpoint is not None:
+        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
             logger.info(
                 f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
                 "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
@@ -419,12 +419,11 @@ def group_texts(examples):
 
     # Training
     if training_args.do_train:
-        if last_checkpoint is not None:
+        checkpoint = None
+        if training_args.resume_from_checkpoint is not None:
+            checkpoint = training_args.resume_from_checkpoint
+        elif last_checkpoint is not None:
             checkpoint = last_checkpoint
-        elif model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path):
-            checkpoint = model_args.model_name_or_path
-        else:
-            checkpoint = None
         train_result = trainer.train(resume_from_checkpoint=checkpoint)
         trainer.save_model()  # Saves the tokenizer too for easy upload
         metrics = train_result.metrics

examples/pytorch/multiple-choice/run_swag.py

@@ -223,7 +223,7 @@ def main():
                 f"Output directory ({training_args.output_dir}) already exists and is not empty. "
                 "Use --overwrite_output_dir to overcome."
             )
-        elif last_checkpoint is not None:
+        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
             logger.info(
                 f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
                 "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
@@ -398,12 +398,11 @@ def compute_metrics(eval_predictions):
 
     # Training
     if training_args.do_train:
-        if last_checkpoint is not None:
+        checkpoint = None
+        if training_args.resume_from_checkpoint is not None:
+            checkpoint = training_args.resume_from_checkpoint
+        elif last_checkpoint is not None:
             checkpoint = last_checkpoint
-        elif os.path.isdir(model_args.model_name_or_path):
-            checkpoint = model_args.model_name_or_path
-        else:
-            checkpoint = None
         train_result = trainer.train(resume_from_checkpoint=checkpoint)
         trainer.save_model()  # Saves the tokenizer too for easy upload
         metrics = train_result.metrics

examples/pytorch/question-answering/run_qa.py

@@ -216,7 +216,7 @@ def main():
                 f"Output directory ({training_args.output_dir}) already exists and is not empty. "
                 "Use --overwrite_output_dir to overcome."
             )
-        elif last_checkpoint is not None:
+        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
             logger.info(
                 f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
                 "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
@@ -557,12 +557,11 @@ def compute_metrics(p: EvalPrediction):
 
     # Training
     if training_args.do_train:
-        if last_checkpoint is not None:
+        checkpoint = None
+        if training_args.resume_from_checkpoint is not None:
+            checkpoint = training_args.resume_from_checkpoint
+        elif last_checkpoint is not None:
             checkpoint = last_checkpoint
-        elif os.path.isdir(model_args.model_name_or_path):
-            checkpoint = model_args.model_name_or_path
-        else:
-            checkpoint = None
         train_result = trainer.train(resume_from_checkpoint=checkpoint)
         trainer.save_model()  # Saves the tokenizer too for easy upload
 

examples/pytorch/question-answering/run_qa_beam_search.py

@@ -215,7 +215,7 @@ def main():
                 f"Output directory ({training_args.output_dir}) already exists and is not empty. "
                 "Use --overwrite_output_dir to overcome."
             )
-        elif last_checkpoint is not None:
+        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
             logger.info(
                 f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
                 "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
@@ -595,12 +595,11 @@ def compute_metrics(p: EvalPrediction):
 
     # Training
     if training_args.do_train:
-        if last_checkpoint is not None:
+        checkpoint = None
+        if training_args.resume_from_checkpoint is not None:
+            checkpoint = training_args.resume_from_checkpoint
+        elif last_checkpoint is not None:
             checkpoint = last_checkpoint
-        elif os.path.isdir(model_args.model_name_or_path):
-            checkpoint = model_args.model_name_or_path
-        else:
-            checkpoint = None
         train_result = trainer.train(resume_from_checkpoint=checkpoint)
         trainer.save_model()  # Saves the tokenizer too for easy upload
 

examples/pytorch/summarization/run_summarization.py

@@ -272,7 +272,7 @@ def main():
                 f"Output directory ({training_args.output_dir}) already exists and is not empty. "
                 "Use --overwrite_output_dir to overcome."
             )
-        elif last_checkpoint is not None:
+        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
             logger.info(
                 f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
                 "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
@@ -520,12 +520,11 @@ def compute_metrics(eval_preds):
 
     # Training
     if training_args.do_train:
-        if last_checkpoint is not None:
+        checkpoint = None
+        if training_args.resume_from_checkpoint is not None:
+            checkpoint = training_args.resume_from_checkpoint
+        elif last_checkpoint is not None:
             checkpoint = last_checkpoint
-        elif os.path.isdir(model_args.model_name_or_path):
-            checkpoint = model_args.model_name_or_path
-        else:
-            checkpoint = None
         train_result = trainer.train(resume_from_checkpoint=checkpoint)
         trainer.save_model()  # Saves the tokenizer too for easy upload
 

examples/pytorch/text-classification/run_glue.py

@@ -196,7 +196,7 @@ def main():
                 f"Output directory ({training_args.output_dir}) already exists and is not empty. "
                 "Use --overwrite_output_dir to overcome."
             )
-        elif last_checkpoint is not None:
+        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
             logger.info(
                 f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
                 "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
@@ -448,14 +448,10 @@ def compute_metrics(p: EvalPrediction):
     # Training
     if training_args.do_train:
         checkpoint = None
-        if last_checkpoint is not None:
+        if training_args.resume_from_checkpoint is not None:
+            checkpoint = training_args.resume_from_checkpoint
+        elif last_checkpoint is not None:
             checkpoint = last_checkpoint
-        elif os.path.isdir(model_args.model_name_or_path):
-            # Check the config from that potential checkpoint has the right number of labels before using it as a
-            # checkpoint.
-            if AutoConfig.from_pretrained(model_args.model_name_or_path).num_labels == num_labels:
-                checkpoint = model_args.model_name_or_path
-
         train_result = trainer.train(resume_from_checkpoint=checkpoint)
         metrics = train_result.metrics
         max_train_samples = (

examples/pytorch/text-classification/run_xnli.py

@@ -335,13 +335,10 @@ def compute_metrics(p: EvalPrediction):
     # Training
     if training_args.do_train:
         checkpoint = None
-        if last_checkpoint is not None:
+        if training_args.resume_from_checkpoint is not None:
+            checkpoint = training_args.resume_from_checkpoint
+        elif last_checkpoint is not None:
             checkpoint = last_checkpoint
-        elif os.path.isdir(model_args.model_name_or_path):
-            # Check the config from that potential checkpoint has the right number of labels before using it as a
-            # checkpoint.
-            if AutoConfig.from_pretrained(model_args.model_name_or_path).num_labels == num_labels:
-                checkpoint = model_args.model_name_or_path
         train_result = trainer.train(resume_from_checkpoint=checkpoint)
         metrics = train_result.metrics
         max_train_samples = (

examples/pytorch/token-classification/run_ner.py

@@ -189,7 +189,7 @@ def main():
                 f"Output directory ({training_args.output_dir}) already exists and is not empty. "
                 "Use --overwrite_output_dir to overcome."
             )
-        elif last_checkpoint is not None:
+        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
             logger.info(
                 f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
                 "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
@@ -437,12 +437,11 @@ def compute_metrics(p):
 
     # Training
     if training_args.do_train:
-        if last_checkpoint is not None:
+        checkpoint = None
+        if training_args.resume_from_checkpoint is not None:
+            checkpoint = training_args.resume_from_checkpoint
+        elif last_checkpoint is not None:
             checkpoint = last_checkpoint
-        elif os.path.isdir(model_args.model_name_or_path):
-            checkpoint = model_args.model_name_or_path
-        else:
-            checkpoint = None
         train_result = trainer.train(resume_from_checkpoint=checkpoint)
         metrics = train_result.metrics
         trainer.save_model()  # Saves the tokenizer too for easy upload

examples/pytorch/translation/run_translation.py

@@ -256,7 +256,7 @@ def main():
                 f"Output directory ({training_args.output_dir}) already exists and is not empty. "
                 "Use --overwrite_output_dir to overcome."
             )
-        elif last_checkpoint is not None:
+        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
             logger.info(
                 f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
                 "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
@@ -512,12 +512,11 @@ def compute_metrics(eval_preds):
 
     # Training
     if training_args.do_train:
-        if last_checkpoint is not None:
+        checkpoint = None
+        if training_args.resume_from_checkpoint is not None:
+            checkpoint = training_args.resume_from_checkpoint
+        elif last_checkpoint is not None:
             checkpoint = last_checkpoint
-        elif os.path.isdir(model_args.model_name_or_path):
-            checkpoint = model_args.model_name_or_path
-        else:
-            checkpoint = None
         train_result = trainer.train(resume_from_checkpoint=checkpoint)
         trainer.save_model()  # Saves the tokenizer too for easy upload
 

src/transformers/training_args.py

@@ -301,6 +301,11 @@ class TrainingArguments:
             :class:`~transformers.Trainer`, it's intended to be used by your training/evaluation scripts instead. See
             the `example scripts <https://github.com/huggingface/transformers/tree/master/examples>`__ for more
             details.
+        resume_from_checkpoint (:obj:`str`, `optional`):
+            The path to a folder with a valid checkpoint for your model. This argument is not directly used by
+            :class:`~transformers.Trainer`, it's intended to be used by your training/evaluation scripts instead. See
+            the `example scripts <https://github.com/huggingface/transformers/tree/master/examples>`__ for more
+            details.
     """
 
     output_dir: str = field(
@@ -531,6 +536,10 @@ class TrainingArguments:
     push_to_hub: bool = field(
         default=False, metadata={"help": "Whether or not to upload the trained model to the model hub after training."}
     )
+    resume_from_checkpoint: Optional[str] = field(
+        default=None,
+        metadata={"help": "The path to a folder with a valid checkpoint for your model."},
+    )
     _n_gpu: int = field(init=False, repr=False, default=-1)
     mp_parameters: str = field(
         default="",