Add losses together

1

Hi

Suppose I have too many losses that need to be added together and I don’t want to do it in the normal way, i.e. L = l1 + l2...

Can I do something like:

loss = torch.autograd.Variable(torch.FloatTensor([0])).cuda()
for i in range(num):
    loss += l[i]

Is this the correct way to do it?

Many Thanks

2

Yes if you make sure that loss is on the same device as the losses.

Why not using the normal way?

3

Hi Simon

Thanks for the reply. cause there are around a hundred losses and I need to select some of them based on a certain criterion. By using the normal way, I have to change the code every time the criterion changes.

4

Okay, that makes sense. In fact you can even start with just plain python number, i.e. loss = 0.