Fix steps bugs in no trainer examples (#24197)

Fix step bugs in no trainer + load checkpoint + grad acc

examples/pytorch/image-classification/run_image_classification_no_trainer.py

@@ -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)

examples/pytorch/image-pretraining/run_mim_no_trainer.py

@@ -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)

examples/pytorch/language-modeling/run_clm_no_trainer.py

@@ -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)

examples/pytorch/language-modeling/run_mlm_no_trainer.py

@@ -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)

examples/pytorch/multiple-choice/run_swag_no_trainer.py

@@ -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)

examples/pytorch/question-answering/run_qa_beam_search_no_trainer.py

@@ -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)

examples/pytorch/question-answering/run_qa_no_trainer.py

@@ -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)

examples/pytorch/semantic-segmentation/run_semantic_segmentation_no_trainer.py

@@ -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)

examples/pytorch/summarization/run_summarization_no_trainer.py

@@ -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)

examples/pytorch/text-classification/run_glue_no_trainer.py

@@ -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()

examples/pytorch/token-classification/run_ner_no_trainer.py

@@ -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()

examples/pytorch/translation/run_translation_no_trainer.py

@@ -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)