Building a mutual attention model from scratch with PyTorch