[losses] migrate npairs multilabel loss

[WindQAQ]

Closes #26. Finally the last one :-)

• Migrate npairs_loss_multilabel from contrib

• Rename npairs_loss_multilabel to npairs_multilabel_loss, which makes sense to addons' naming style

• Modify input spec from

  • labels (list of sparse tensor)
  • embedding_anchor
  • embedding_posititve

  to

  • y_true (either tensor or sparse tensor)
  • y_pred (similarity matrix, which is the multiplication of anchor and positive)

[facaiy]

Looks good, Tzu-wei!

By the way, can we provide a helper method which calculates similarity_matrix? It might be user-friendly for tf 1.x users. What do you think?

[facaiy]

@Squadrick Hi, Dheeraj. I think the PR looks good. Do you have time to double check it? Thanks :-)

[WindQAQ]

Looks good, Tzu-wei!

By the way, can we provide a helper method which calculates similarity_matrix? It might be user-friendly for tf 1.x users. What do you think?

Seems good, but it will be a one line function like tf.matmul(a, b).

[facaiy]

@WindQAQ Oh, you're right, forget it

[Squadrick]

Maybe add a test to check that the new loss is compatible with the Keras Model API? It would be great if we add it under utils and every loss has to use the check. What do y'all think?

[facaiy]

Like layer_test or https://github.com/tensorflow/tensorflow/blob/0197a2d8a3070af763cb67227835ee63df095e6d/tensorflow/python/keras/losses_test.py#L68 ? I have no objection. How about filing an issue and finishing it in next PR?

[facaiy]

I'd like to merge it at first, thank you, Tzu-Wei!