Bug for evaluating the contrastive_excitation backprop with pointing game

[wjNam]

Hi, thank you for your excellent package.

I found some bug that some results of CEB which have nan values get hit points.
(this occurred when I changed return value as
def gradient_to_contrastive_excitation_backprop_saliency(x):
return torch.sum(x.grad, 1, keepdim=True)
Is it related to the final result?
I think inf value counted as the highest value.

Please check.