Incorrect use of torch.nn.utils.clip_grad_norm

[zazyzaya]

In both training examples clip_grad_norm() is called after opt.step() so it isn't doing anything. It should either be removed or moved before the optimizer updates the model components