Different gradients under the same condition when generating adversarial examples

vision
2

What happens when you zero_grad before the forward? I.e.,
res101.zero_grad()
loss1= res101(normed_input).sum()