HierarchicalSampler is not passed as a batch_sampler parameter

[ManuelZ]

I'm trying to use the MetricLossOnly trainer with the HierarchicalSampler, but it's inheriting from Sampler instead of from BatchSampler and due to this, the following block isn't executed.

pytorch-metric-learning/src/pytorch_metric_learning/utils/common_functions.py

Lines 176 to 183 in c835099

	if isinstance(sampler, torch.utils.data.BatchSampler):
	return torch.utils.data.DataLoader(
	dataset,
	batch_sampler=sampler,
	num_workers=num_workers,
	collate_fn=collate_fn,
	pin_memory=False,
	)

And the training doesn't work because instead of getting a 4D tensor like b,c,h,w, I get b,b,c,h,w and an error.

I haven't digged further, but so far, the training seems to be working if I force the HierarchicalSampler to inherit from BatchSampler.

Why did you change it from BatchSampler to Sampler?

Thanks for this fantastic library!

[KevinMusgrave]

Yeah, it looks like it should extend BatchSampler as you say. If you'd like you can open a pull request with the fix.