get_grad( ) function returns NaN

[bussfromspace]

Hi,
I was trying to experiment with partial information attack with my ImageNet samples but get_grad() function on line 205 returns nan for both loss and gradient estimate. Therefore, it always gives "Did not converge" error.
Query-limited scenario works quite well, i think the problem is inside partial_info_loss.

[Xuejing0203]

I met same problem, do you remember how you solved it