chore: correct update_step and correct gradient_accumulation_steps (#26068)

examples/pytorch/image-classification/run_image_classification_no_trainer.py

@@ -477,8 +477,8 @@ def main():
             # 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)
+            completed_steps = resume_step // args.gradient_accumulation_steps
             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)

examples/pytorch/image-pretraining/run_mim_no_trainer.py

@@ -701,8 +701,8 @@ def main():
             # 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_steps
+            resume_step -= starting_epoch * len(train_dataloader)
 
     # update the progress_bar if load from checkpoint
     progress_bar.update(completed_steps)

examples/pytorch/language-modeling/run_mlm_no_trainer.py

@@ -636,8 +636,8 @@ def main():
             # 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_steps
+            resume_step -= starting_epoch * len(train_dataloader)
 
     # update the progress_bar if load from checkpoint
     progress_bar.update(completed_steps)

examples/pytorch/multiple-choice/run_swag_no_trainer.py

@@ -583,8 +583,8 @@ def main():
             # 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)
+            completed_steps = resume_step // args.gradient_accumulation_steps
             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)

examples/pytorch/question-answering/run_qa_beam_search_no_trainer.py

@@ -820,8 +820,8 @@ def main():
             # 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)
+            completed_steps = resume_step // args.gradient_accumulation_steps
             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)

examples/pytorch/question-answering/run_qa_no_trainer.py

@@ -848,10 +848,11 @@ 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)
+            completed_steps = resume_step // args.gradient_accumulation_steps
             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)

examples/pytorch/semantic-segmentation/run_semantic_segmentation_no_trainer.py

@@ -581,8 +581,8 @@ def main():
             # 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)
+            completed_steps = resume_step // args.gradient_accumulation_steps
             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)

examples/pytorch/summarization/run_summarization_no_trainer.py

@@ -652,8 +652,8 @@ def main():
             # 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)
+            completed_steps = resume_step // args.gradient_accumulation_steps
             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)

examples/pytorch/text-classification/run_glue_no_trainer.py

@@ -530,8 +530,8 @@ def main():
             # 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)
+            completed_steps = resume_step // args.gradient_accumulation_steps
             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)

examples/pytorch/token-classification/run_ner_no_trainer.py

@@ -690,8 +690,8 @@ def main():
             # 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)
+            completed_steps = resume_step // args.gradient_accumulation_steps
             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)

examples/pytorch/translation/run_translation_no_trainer.py

@@ -633,8 +633,8 @@ def main():
             # 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)
+            completed_steps = resume_step // args.gradient_accumulation_steps
             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)