Loss function class

[artidoro]

In Pytorch loss functions in torch.nn work as follows:

loss = NLLLoss(SOME PARAMETERS IF NEEDED)
prediction = model.forward(input)
output = loss(prediction, target)

We should match the behavior where once initialized the loss can be applied to the prediction and the target tensors.

[interesaaat]

I am thinking about providing an additional method to the loss class whereby you can specify parameters for the loss function you want to use and a lambda function is returned (instead of an object). Do you think that this could work?

This is the only way we can actually implement loss(prediction, target) in C#

[artidoro]

Yes, I think that should work!