riaz.somc/transformers
1
0
Fork 0
mirror of https://github.com/huggingface/transformers.git synced 2025-07-24 23:08:57 +06:00
Code Issues Actions 35 Packages Projects Releases Wiki Activity

Fix steps bugs in no trainer examples (#24197)

Browse Source 
Fix step bugs in no trainer + load checkpoint + grad acc
This commit is contained in:
Ethan 2023-06-12 23:49:55 +08:00 committed by GitHub
parent 08ae37c820
commit f7d80cb3d2
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
12 changed files with 34 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
examples/pytorch/image-classification/run_image_classification_no_trainer.py
View File

@ -453,10 +453,11 @@ def main():
resume_step = None
completed_steps = starting_epoch * num_update_steps_per_epoch
else:
resume_step = int(training_difference.replace("step_", ""))
# need to multiply `gradient_accumulation_steps` to reflect real steps
resume_step = int(training_difference.replace("step_", "")) * args.gradient_accumulation_steps
starting_epoch = resume_step // len(train_dataloader)
resume_step -= starting_epoch * len(train_dataloader)
completed_steps = resume_step
completed_steps = resume_step // args.gradient_accumulation_step
# update the progress_bar if load from checkpoint
progress_bar.update(completed_steps)

2
examples/pytorch/image-pretraining/run_mim_no_trainer.py
View File

@ -666,7 +666,7 @@ def main():
resume_step = int(training_difference.replace("step_", "")) * args.gradient_accumulation_steps
starting_epoch = resume_step // len(train_dataloader)
resume_step -= starting_epoch * len(train_dataloader)
completed_steps = resume_step
completed_steps = resume_step // args.gradient_accumulation_steps
# update the progress_bar if load from checkpoint
progress_bar.update(completed_steps)

2
examples/pytorch/language-modeling/run_clm_no_trainer.py
View File

@ -572,7 +572,7 @@ def main():
resume_step = int(training_difference.replace("step_", "")) * args.gradient_accumulation_steps
starting_epoch = resume_step // len(train_dataloader)
resume_step -= starting_epoch * len(train_dataloader)
completed_steps = resume_step
completed_steps = resume_step // args.gradient_accumulation_steps
# update the progress_bar if load from checkpoint
progress_bar.update(completed_steps)

2
examples/pytorch/language-modeling/run_mlm_no_trainer.py
View File

@ -616,7 +616,7 @@ def main():
resume_step = int(training_difference.replace("step_", "")) * args.gradient_accumulation_steps
starting_epoch = resume_step // len(train_dataloader)
resume_step -= starting_epoch * len(train_dataloader)
completed_steps = resume_step
completed_steps = resume_step // args.gradient_accumulation_steps
# update the progress_bar if load from checkpoint
progress_bar.update(completed_steps)

5
examples/pytorch/multiple-choice/run_swag_no_trainer.py
View File

@ -559,10 +559,11 @@ def main():
resume_step = None
completed_steps = starting_epoch * num_update_steps_per_epoch
else:
resume_step = int(training_difference.replace("step_", ""))
# need to multiply `gradient_accumulation_steps` to reflect real steps
resume_step = int(training_difference.replace("step_", "")) * args.gradient_accumulation_steps
starting_epoch = resume_step // len(train_dataloader)
resume_step -= starting_epoch * len(train_dataloader)
completed_steps = resume_step
completed_steps = resume_step // args.gradient_accumulation_stepp
# update the progress_bar if load from checkpoint
progress_bar.update(completed_steps)

5
examples/pytorch/question-answering/run_qa_beam_search_no_trainer.py
View File

@ -811,10 +811,11 @@ def main():
resume_step = None
completed_steps = starting_epoch * num_update_steps_per_epoch
else:
resume_step = int(training_difference.replace("step_", ""))
# need to multiply `gradient_accumulation_steps` to reflect real steps
resume_step = int(training_difference.replace("step_", "")) * args.gradient_accumulation_steps
starting_epoch = resume_step // len(train_dataloader)
resume_step -= starting_epoch * len(train_dataloader)
completed_steps = resume_step
completed_steps = resume_step // args.gradient_accumulation_stepp
# update the progress_bar if load from checkpoint
progress_bar.update(completed_steps)

2
examples/pytorch/question-answering/run_qa_no_trainer.py
View File

@ -830,7 +830,7 @@ def main():
resume_step = int(training_difference.replace("step_", ""))
starting_epoch = resume_step // len(train_dataloader)
resume_step -= starting_epoch * len(train_dataloader)
completed_steps = resume_step
completed_steps = resume_step // args.gradient_accumulation_stepp
# update the progress_bar if load from checkpoint
progress_bar.update(completed_steps)

5
examples/pytorch/semantic-segmentation/run_semantic_segmentation_no_trainer.py
View File

@ -556,10 +556,11 @@ def main():
resume_step = None
completed_steps = starting_epoch * num_update_steps_per_epoch
else:
resume_step = int(training_difference.replace("step_", ""))
# need to multiply `gradient_accumulation_steps` to reflect real steps
resume_step = int(training_difference.replace("step_", "")) * args.gradient_accumulation_steps
starting_epoch = resume_step // len(train_dataloader)
resume_step -= starting_epoch * len(train_dataloader)
completed_steps = resume_step
completed_steps = resume_step // args.gradient_accumulation_stepp
# update the progress_bar if load from checkpoint
progress_bar.update(completed_steps)

5
examples/pytorch/summarization/run_summarization_no_trainer.py
View File

@ -628,10 +628,11 @@ def main():
resume_step = None
completed_steps = starting_epoch * num_update_steps_per_epoch
else:
resume_step = int(training_difference.replace("step_", ""))
# need to multiply `gradient_accumulation_steps` to reflect real steps
resume_step = int(training_difference.replace("step_", "")) * args.gradient_accumulation_steps
starting_epoch = resume_step // len(train_dataloader)
resume_step -= starting_epoch * len(train_dataloader)
completed_steps = resume_step
completed_steps = resume_step // args.gradient_accumulation_stepp
# update the progress_bar if load from checkpoint
progress_bar.update(completed_steps)

8
examples/pytorch/text-classification/run_glue_no_trainer.py
View File

@ -501,10 +501,16 @@ def main():
if "epoch" in training_difference:
starting_epoch = int(training_difference.replace("epoch_", "")) + 1
resume_step = None
completed_steps = starting_epoch * num_update_steps_per_epoch
else:
resume_step = int(training_difference.replace("step_", ""))
# need to multiply `gradient_accumulation_steps` to reflect real steps
resume_step = int(training_difference.replace("step_", "")) * args.gradient_accumulation_steps
starting_epoch = resume_step // len(train_dataloader)
resume_step -= starting_epoch * len(train_dataloader)
completed_steps = resume_step // args.gradient_accumulation_step
# update the progress_bar if load from checkpoint
progress_bar.update(completed_steps)
for epoch in range(starting_epoch, args.num_train_epochs):
model.train()

8
examples/pytorch/token-classification/run_ner_no_trainer.py
View File

@ -659,10 +659,16 @@ def main():
if "epoch" in training_difference:
starting_epoch = int(training_difference.replace("epoch_", "")) + 1
resume_step = None
completed_steps = starting_epoch * num_update_steps_per_epoch
else:
resume_step = int(training_difference.replace("step_", ""))
# need to multiply `gradient_accumulation_steps` to reflect real steps
resume_step = int(training_difference.replace("step_", "")) * args.gradient_accumulation_steps
starting_epoch = resume_step // len(train_dataloader)
resume_step -= starting_epoch * len(train_dataloader)
completed_steps = resume_step // args.gradient_accumulation_stepp
# update the progress_bar if load from checkpoint
progress_bar.update(completed_steps)
for epoch in range(starting_epoch, args.num_train_epochs):
model.train()

2
examples/pytorch/translation/run_translation_no_trainer.py
View File

@ -613,7 +613,7 @@ def main():
resume_step = int(training_difference.replace("step_", "")) * args.gradient_accumulation_steps
starting_epoch = resume_step // len(train_dataloader)
resume_step -= starting_epoch * len(train_dataloader)
completed_steps = resume_step
completed_steps = resume_step // args.gradient_accumulation_stepp
# update the progress_bar if load from checkpoint
progress_bar.update(completed_steps)