Gradients exist but weights not updating

Hi,

When you get the parameters of your net, it does not clone the tensors. So in your case, before and after contain the same tensors. So when the optimizer update the weights in place, it updates both your lists. You can try and change one weight by hand, they will still remain the same.