[feat] Implementing threshold consistent margin loss

[ir2718]

Hi,

this PR implements the loss presented in the paper Threshold-Consistent Margin Loss for Open-World Deep Metric Learning. The loss is combined with a base loss provided by the user. The default hyperparameters are chosen using the values provided in the section 5.2.

Aside from the implementation, I also added a section in the docs and a test case.

[KevinMusgrave]

Thank you @ir2718!

I wonder if we should get rid of the base_loss argument, and let the user add the two together themselves. For example, if the base loss is ContrastiveLoss then:

loss_fn = ThresholdConsistentMarginLoss()
base_loss_fn = ContrastiveLoss()

# when computing loss
loss = loss_fn(embeddings, labels) + base_loss_fn(embeddings, labels)

And if a user really wants to combine them into a single function call, they can wrap it in a function:

def total_loss():
    return  loss_fn(embeddings, labels) + base_loss_fn(embeddings, labels)

Or use the MultipleLosses wrapper:

from pytorch_metric_learning.losses import MultipleLosses
loss_fn = MultipleLosses([ContrastiveLoss(), ThresholdConsistentMarginLoss()])

[ir2718]

Oh, I wasn't aware MultipleLosses existed. I removed the base loss as requested and updated the docs accordingly.

[KevinMusgrave]

Thanks @ir2718!