Grad record scale

[jacenfox]

It's a bug fixed version.

[MathGaron]

This is already implemented in the framework. Check the cat_dog exemple :
in your network init :
Basically you pass a lambda function that takes the gradient as input and outputs whatever
and pass a tag name so you can retrieve it (fc1 in this case)
self.hook_fc1 = self.hook_generator(lambda x: float(torch.sqrt(torch.sum(x ** 2))), "fc1")
in your network forward :
if self.training: x.register_hook(self.hook_fc1)
and to retrieve the gradient in your callback:
grad_log = state.model.grad_data["fc1"]

I am sorry about that, I should really take the time to document it properly....