Which loss in pytorch is similar SoftmaxWithLoss in caffe?

1

Hello all, In caffe I used the SoftmaxWithLoss for multiple class segmentation problem

(Caffe) block (n) → BatchNorm → ReLU → SoftmaxWithLoss

Which loss in pytorch should I used to achieve similar result in caffe? Thanks

2

I think that would be

import torch.nn.functional as F
F.cross_entropy()

or the equivalent (object-oriented API)

torch.nn.CrossEntropyLoss

These take the logits as inputs and compute the log softmax and pass it to the neg log likelihood loss (/multinomial logistic loss) function internally.

1 Like