Calculate gradient in Partial information setting

[liuzhidanhhh]

Hi，
When I run this code to make an adversarial examples in Partial information setting, I an confused by a problem:
the function of partial_info_loss has no difference with the function standard_loss.
The code of line 125 to line 130 doesn't play an role on selecting topK logit and noise, because good_image is euqal to range(batch_size) . When i delete those lines, I got the same attack result.
If the function of partial_info_loss did not select topK logit, how this algorithm calculate gradient in this setting？