Classification references does not work without distributed setup

[pmeier]

If you don't set the respective env vars

vision/references/classification/utils.py

Lines 255 to 258 in d5bd8b7

	else:
	print("Not using distributed mode")
	args.distributed = False
	return

training will not be distributed and in turn the backend will not be initialized. However, during evaluation we check

vision/references/classification/train.py

Line 88 in d5bd8b7

and torch.distributed.get_rank() == 0

unguarded, which then fails with

RuntimeError: Default process group has not been initialized, please make sure to call init_process_group.

cc @datumbox

[pmeier]

Same for segmentation:

vision/references/segmentation/train.py

Line 84 in cac4e22

and torch.distributed.get_rank() == 0

[YosuaMichael]

I think this case is implicitly guarded on

vision/references/classification/train.py

Line 87 in d5bd8b7

and len(data_loader.dataset) != num_processed_samples

since len(data_loader.dataset) != num_processed_samples shouldn't be true on non-distributed setting.

Do you get the error during non-distributed training @pmeier ?