SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Update no trainer examples for QA and Semantic Segmentation (#18474)

Browse Source 
* swag_no_trainer updated for with gather_metrics

* Removed unused variable samples_seen

* updated examples with gather_for_metrics
This commit is contained in:
Kian Sierra McGettigan
2022-08-04 19:22:19 +02:00
committed by GitHub
parent d2704c4143
commit 0bf1e1aca4
3 changed files with 17 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
examples/pytorch/question-answering/run_qa_beam_search_no_trainer.py
View File

@@ -698,7 +698,7 @@ def main():
step = 0 step = 0
# create a numpy array and fill it with -100. # create a numpy array and fill it with -100.
logits_concat = np.full((len(dataset), max_len), -100, dtype=np.float32) logits_concat = np.full((len(dataset), max_len), -100, dtype=np.float32)
# Now since we have create an array now we will populate it with the outputs gathered using accelerator.gather # Now since we have create an array now we will populate it with the outputs gathered using accelerator.gather_for_metrics
for i, output_logit in enumerate(start_or_end_logits): # populate columns for i, output_logit in enumerate(start_or_end_logits): # populate columns
# We have to fill it such that we have to take the whole tensor and replace it on the newly created array # We have to fill it such that we have to take the whole tensor and replace it on the newly created array
# And after every iteration we have to change the step # And after every iteration we have to change the step
@@ -876,11 +876,11 @@ def main():
end_top_index = accelerator.pad_across_processes(end_top_index, dim=1, pad_index=-100) end_top_index = accelerator.pad_across_processes(end_top_index, dim=1, pad_index=-100)
cls_logits = accelerator.pad_across_processes(cls_logits, dim=1, pad_index=-100) cls_logits = accelerator.pad_across_processes(cls_logits, dim=1, pad_index=-100)
all_start_top_log_probs.append(accelerator.gather(start_top_log_probs).cpu().numpy()) all_start_top_log_probs.append(accelerator.gather_for_metrics(start_top_log_probs).cpu().numpy())
all_start_top_index.append(accelerator.gather(start_top_index).cpu().numpy()) all_start_top_index.append(accelerator.gather_for_metrics(start_top_index).cpu().numpy())
all_end_top_log_probs.append(accelerator.gather(end_top_log_probs).cpu().numpy()) all_end_top_log_probs.append(accelerator.gather_for_metrics(end_top_log_probs).cpu().numpy())
all_end_top_index.append(accelerator.gather(end_top_index).cpu().numpy()) all_end_top_index.append(accelerator.gather_for_metrics(end_top_index).cpu().numpy())
all_cls_logits.append(accelerator.gather(cls_logits).cpu().numpy()) all_cls_logits.append(accelerator.gather_for_metrics(cls_logits).cpu().numpy())
max_len = max([x.shape[1] for x in all_end_top_log_probs]) # Get the max_length of the tensor max_len = max([x.shape[1] for x in all_end_top_log_probs]) # Get the max_length of the tensor
@@ -936,11 +936,11 @@ def main():
end_top_index = accelerator.pad_across_processes(end_top_index, dim=1, pad_index=-100) end_top_index = accelerator.pad_across_processes(end_top_index, dim=1, pad_index=-100)
cls_logits = accelerator.pad_across_processes(cls_logits, dim=1, pad_index=-100) cls_logits = accelerator.pad_across_processes(cls_logits, dim=1, pad_index=-100)
all_start_top_log_probs.append(accelerator.gather(start_top_log_probs).cpu().numpy()) all_start_top_log_probs.append(accelerator.gather_for_metrics(start_top_log_probs).cpu().numpy())
all_start_top_index.append(accelerator.gather(start_top_index).cpu().numpy()) all_start_top_index.append(accelerator.gather_for_metrics(start_top_index).cpu().numpy())
all_end_top_log_probs.append(accelerator.gather(end_top_log_probs).cpu().numpy()) all_end_top_log_probs.append(accelerator.gather_for_metrics(end_top_log_probs).cpu().numpy())
all_end_top_index.append(accelerator.gather(end_top_index).cpu().numpy()) all_end_top_index.append(accelerator.gather_for_metrics(end_top_index).cpu().numpy())
all_cls_logits.append(accelerator.gather(cls_logits).cpu().numpy()) all_cls_logits.append(accelerator.gather_for_metrics(cls_logits).cpu().numpy())
max_len = max([x.shape[1] for x in all_end_top_log_probs]) # Get the max_length of the tensor max_len = max([x.shape[1] for x in all_end_top_log_probs]) # Get the max_length of the tensor

10
examples/pytorch/question-answering/run_qa_no_trainer.py
View File

@@ -715,7 +715,7 @@ def main():
step = 0 step = 0
# create a numpy array and fill it with -100. # create a numpy array and fill it with -100.
logits_concat = np.full((len(dataset), max_len), -100, dtype=np.float64) logits_concat = np.full((len(dataset), max_len), -100, dtype=np.float64)
# Now since we have create an array now we will populate it with the outputs gathered using accelerator.gather # Now since we have create an array now we will populate it with the outputs gathered using accelerator.gather_for_metrics
for i, output_logit in enumerate(start_or_end_logits): # populate columns for i, output_logit in enumerate(start_or_end_logits): # populate columns
# We have to fill it such that we have to take the whole tensor and replace it on the newly created array # We have to fill it such that we have to take the whole tensor and replace it on the newly created array
# And after every iteration we have to change the step # And after every iteration we have to change the step
@@ -901,8 +901,8 @@ def main():
start_logits = accelerator.pad_across_processes(start_logits, dim=1, pad_index=-100) start_logits = accelerator.pad_across_processes(start_logits, dim=1, pad_index=-100)
end_logits = accelerator.pad_across_processes(end_logits, dim=1, pad_index=-100) end_logits = accelerator.pad_across_processes(end_logits, dim=1, pad_index=-100)
all_start_logits.append(accelerator.gather(start_logits).cpu().numpy()) all_start_logits.append(accelerator.gather_for_metrics(start_logits).cpu().numpy())
all_end_logits.append(accelerator.gather(end_logits).cpu().numpy()) all_end_logits.append(accelerator.gather_for_metrics(end_logits).cpu().numpy())
max_len = max([x.shape[1] for x in all_start_logits]) # Get the max_length of the tensor max_len = max([x.shape[1] for x in all_start_logits]) # Get the max_length of the tensor
@@ -940,8 +940,8 @@ def main():
start_logits = accelerator.pad_across_processes(start_logits, dim=1, pad_index=-100) start_logits = accelerator.pad_across_processes(start_logits, dim=1, pad_index=-100)
end_logits = accelerator.pad_across_processes(end_logits, dim=1, pad_index=-100) end_logits = accelerator.pad_across_processes(end_logits, dim=1, pad_index=-100)
all_start_logits.append(accelerator.gather(start_logits).cpu().numpy()) all_start_logits.append(accelerator.gather_for_metrics(start_logits).cpu().numpy())
all_end_logits.append(accelerator.gather(end_logits).cpu().numpy()) all_end_logits.append(accelerator.gather_for_metrics(end_logits).cpu().numpy())
max_len = max([x.shape[1] for x in all_start_logits]) # Get the max_length of the tensor max_len = max([x.shape[1] for x in all_start_logits]) # Get the max_length of the tensor
# concatenate the numpy array # concatenate the numpy array

11
examples/pytorch/semantic-segmentation/run_semantic_segmentation_no_trainer.py
View File

@@ -605,7 +605,6 @@ def main():
logger.info("***** Running evaluation *****") logger.info("***** Running evaluation *****")
model.eval() model.eval()
samples_seen = 0
for step, batch in enumerate(tqdm(eval_dataloader, disable=not accelerator.is_local_main_process)): for step, batch in enumerate(tqdm(eval_dataloader, disable=not accelerator.is_local_main_process)):
with torch.no_grad(): with torch.no_grad():
outputs = model(**batch) outputs = model(**batch)
@@ -615,15 +614,7 @@ def main():
) )
predictions = upsampled_logits.argmax(dim=1) predictions = upsampled_logits.argmax(dim=1)
predictions, references = accelerator.gather((predictions, batch["labels"])) predictions, references = accelerator.gather_for_metrics((predictions, batch["labels"]))
# If we are in a multiprocess environment, the last batch has duplicates
if accelerator.num_processes > 1:
if step == len(eval_dataloader) - 1:
predictions = predictions[: len(eval_dataloader.dataset) - samples_seen]
references = references[: len(eval_dataloader.dataset) - samples_seen]
else:
samples_seen += references.shape[0]
metric.add_batch( metric.add_batch(
predictions=predictions, predictions=predictions,