Loss function for reconstruction and its last layer

I am trying to generate a colored image using the traditional encoder-decoder model. Since the pixels are normalized in the range [0,1], it makes sense to have a sigmoid in the last layer of the decoder. Now, for the same, if I have the normalized ground truth image, what should be my loss function? nn.MSE is one option, but I am not keen on using it (something to do with my problem statement). Is there another loss function criterion that I can use?

You could try to use e.g. nn.BCELoss, if it would fit your use case. :wink: