MPerClassSampler: 'numpy.int32' object is not iterable

[rheum]

Hey all. I believe I have found a bug with MPerClassSampler. Using a dataloader which makes use of that sampler throws an exception. As far as I can tell, my HTR code meets all documented preconditions of MPerClassSampler.

My setup:
pytorch_metric_learning.version == "2.1.0"
torch.version == '1.13.1+cu116'

How to reproduce:

from torch.utils.data import TensorDataset, DataLoader
from pytorch_metric_learning.samplers import MPerClassSampler

ds = TensorDataset(torch.Tensor([[1], [2], [3]]), torch.Tensor([0, 0, 1]))
batch_sampler = MPerClassSampler(labels=[0, 0, 1], m=1, batch_size=2)
train_loader = DataLoader(ds, num_workers=0, batch_sampler=batch_sampler)
next(iter(train_loader))

Traceback (most recent call last):
  File "...\AppData\Roaming\Python\Python310\site-packages\IPython\core\interactiveshell.py", line 3460, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "<ipython-input-11-0fd2476fd2ff>", line 1, in <module>
    next(iter(train_loader))
  File "...\AppData\Roaming\Python\Python310\site-packages\torch\utils\data\dataloader.py", line 628, in __next__
    data = self._next_data()
  File ...\AppData\Roaming\Python\Python310\site-packages\torch\utils\data\dataloader.py", line 671, in _next_data
    data = self._dataset_fetcher.fetch(index)  # may raise StopIteration
  File "...\AppData\Roaming\Python\Python310\site-packages\torch\utils\data\_utils\fetch.py", line 58, in fetch
    data = [self.dataset[idx] for idx in possibly_batched_index]
TypeError: 'numpy.int32' object is not iterable

[KevinMusgrave]

Can you try changing batch_sampler=batch_sampler to sampler=batch_sampler ?

[rheum]

Thanks for the quick response!
While there's no longer an exception, the batch_size doesn't seem to work:

[KevinMusgrave]

You have to pass in batch_size=2 into the DataLoader as well.

The reason is that MPerClassSampler is a sampler not a batch_sampler, so the batch size is actually determined by DataLoader.

The batch_size argument for MPerClassSampler is just for internal checks.

It would make more sense for it to be a batch_sampler, then you wouldn't have to specify batch_size in 2 places. I think this has come up a few times, so I've made a separate issue for fixing this: #612

[rheum]

Thank you. Turns out I was distracted by the explicit mention of batch_sampler in the docs for HierarchicalSampler, whereas for MPerClassSampler there is no information on how it is used. Maybe as a quick-fix, this could be clarified in the doc. I'll create a PR later today.