The following question is not a duplicate of How to apply layer-wise learning rate in Pytorch? because this question aims at freezing a subset of a tensor from training rather than the entire layer.
I am trying out a PyTorch implementation of Lottery Ticket Hypothesis.
For that, I want to freeze the weights in a model that are zero. Is the following a correct way to implement it?
for name, p in model.named_parameters():
if 'weight' in name:
tensor = p.data.cpu().numpy()
grad_tensor = p.grad.data.cpu().numpy()
grad_tensor = np.where(tensor == 0, 0, grad_tensor)
p.grad.data = torch.from_numpy(grad_tensor).to(device)
What you have seems like it would work provided you did it after loss.backward() and before optimizer.step() (referring to the common usage for these variable names). That said, it seems a bit convoluted. Also, if your weights are floating point values then comparing them to exactly zero is probably a bad idea, we could introduce an epsilon to account for this. IMO the following is a little cleaner than the solution you proposed:
# locate zero-value weights before training loop
EPS = 1e-6
locked_masks = {n: torch.abs(w) < EPS for n, w in model.named_parameters() if n.endswith('weight')}
...
for ... #training loop
...
optimizer.zero_grad()
loss.backward()
# zero the gradients of interest
for n, w in model.named_parameters():
if w.grad is not None and n in locked_masks:
w.grad[locked_masks[n]] = 0
optimizer.step()