agil27 / tcav_pytorch Goto Github PK

#2 HI, The link you shared for the pre-training model is not working. Can you share it again? Thank you!

• [ ]

The link to download the model is invalid, can you upload again?

Hello agil27;

Thank you for your implementation, this really helpful as I'm coding with pytorch.

I have one question:

• Please, why do you compute backward rather than grad (from autograd) in the function:
  def generate_gradients(self, c, layer_name): activation = self.intermediate_activations[layer_name] activation.register_hook(self.save_gradient) logit = self.output[:, c] logit.backward(torch.ones_like(logit), retain_graph=True) # gradients = grad(logit, activation, retain_graph=True)[0] # gradients = gradients.cpu().detach().numpy() gradients = self.gradients.cpu().detach().numpy() return gradients<br class="Apple-interchange-newline">

the grad function is commented with "#"

Thank you very much,
Habib