get_gradients loss is the output of the NN

[sulantha2006]

This may be a more of a theoretical question. But, when calculating the saliency maps and the guided back prop, when calculating the loss to get the gradients, the output is used (of the class).
Is it how the loss needs to be calculated for saliency and guided back prop? Thank you.

[Max-ChenFei]

I have the same questions. I see some implementations of saliency map use loss while here just use the output of the model. Thank you!