Allow frozen parameters with requires_grad=False

[andrewli77]

Hi Lucas, thanks for open-sourcing your code for this project. I think it would be nice to be able to manually freeze some network parameters (I was doing this in the context of transfer learning). Currently, the gradient norm is computed over parameters with requires_grad=False as well and p.grad is None.

[lcswillems]

Hi @andrewli77, Thank you for your PR! How would you define the value of requires_grad?

[andrewli77]

Hi Lucas, sorry for the late reply. requires_grad is an attribute for every parameter of nn.Module which defaults to true. My code sets it by subclassing ACModel and running

for param in self.rnn.parameters():
         param.requires_grad = False

[lcswillems]

Hi @andrewli77,

I completely forgot about your PR. I am wondering if your modifications change something.

Your first modification:

grad_norm = sum(p.grad.data.norm(2).item() ** 2 for p in self.acmodel.parameters() if p.requires_grad) ** 0.5

If p.requires_grad is false, then p.grad.data would be an array of zeros, right? In this case, removing if p.requires_grad would not change anything right?

Your second modification:

torch.nn.utils.clip_grad_norm_([p for p in self.acmodel.parameters() if p.requires_grad], self.max_grad_norm)

If p.requires_grad is false, then p.grad.data would be an array of zeros and therefore clipping would change nothing.