Your code is quite hard to read since you haven’t formatted it properly.
From the first look it seems these parts might be wrong:
(‘output’, nn.Softmax(dim=0))])).cuda()
criterion = nn.CrossEntropyLoss()
nn.CrossEntropyLoss expects raw logits, so remove the Softmax layer.
scores=max(scores)
loss = criterion(scores, targets)
I don’t know why max is used here instead of passing the logits to the criterion directly.