Question about backward hook

[OValery16]

Thank you for work and sharing this code.

I was wondering the meaning of new_grad = torch.ones_like(grad) in the following code in universal.py.

    def _backward_hook(self, grad):
        self.out_mask += grad.data.abs().sum(2, keepdim=True).sum(3, keepdim=True).cpu().sum(0, keepdim=True).view(-1)
        new_grad = torch.ones_like(grad)
        return new_grad