Fix kwargs not passed to SubCenterArcFaceLoss

[marcpaga]

**kwargs are not passed into __init__ of SubCenterArcFaceLoss.

This leads to problems, for example with reducers, which have no effect when passed.

Example bug:

arcfaceloss = SubCenterArcFaceLoss(
    num_classes=10,
    sub_centers=3,
    embedding_size=64,
    reducer=DoNothingReducer(),
)

emb = torch.randn(4, 64)
print(arcfaceloss(emb, torch.arange(4)))

# without fix, we get the mean of the losses, which is the default reducer, despite passing `DoNothingReducer` to the `init`.
> tensor(39.9293, grad_fn=<MeanBackward0>)

# with fix, the reducer is properly passed and is properly applied.
> {'loss': {'losses': tensor([53.9860, 42.8221, 38.8233, 62.1092], grad_fn=<MulBackward0>), 'indices': tensor([0, 1, 2, 3]), 'reduction_type': 'element'}}

The fix includes the use of deepcopy to ensure no reference problems occur when deleting the keys form **kwargs, which is necessary otherwise duplicate keys are passed into __init__. But perhaps it can be done more elegantly.

[KevinMusgrave]

Thanks! Seems reasonable. Can you add your example to the tests here: https://github.com/KevinMusgrave/pytorch-metric-learning/blob/master/tests/losses/test_subcenter_arcface_loss.py

Something like this should be sufficient I think:

from pytorch_metric_learning.reducers import DoNothingReducer


def test_reducer(self):
  arcfaceloss = SubCenterArcFaceLoss(
      num_classes=10,
      sub_centers=3,
      embedding_size=64,
      reducer=DoNothingReducer(),
  )
  
  emb = torch.randn(4, 64)
  result = arcfaceloss(emb, torch.arange(4))
  self.assertTrue(isinstance(result, dict))

[marcpaga]

Hi @KevinMusgrave,
I have added the unittest, with also a shape check based on the batch dim.
Cheers!

[KevinMusgrave]

Thanks!