How to exclude loss computation on certain tensors in PyTorch