Why does the first term in the loss_robust equal F.log_softmax(model(x_adv))?

[changzhang777]

Why does the first term in the loss_robust equal F.log_softmax(model(x_adv))? I have tryed to generate adv samples by noraml PGD, and set the loss to the trades's loss. But I cannot understand why it is and I guess I prefer to use criterion_kl(F.softmax(model(x_adv)), F.softmax(model(x_natural))). Are there some nice people who can answer my question? I would appreciate it.

[amirhagai]

As far as I understand they try to minimize the cross entropy loss, but in torch this loss expect to get one hot vector.
see equation 1 at https://en.wikipedia.org/wiki/Cross_entropy

[changzhang777]

I think you mean that it is uesd to keep consistent with log_softmaxed vector which is in the cross entropy when optimizing. Thx a lot.

[pingguokiller]

Why does the first term in the loss_robust equal F.log_softmax(model(x_adv))? I have tryed to generate adv samples by noraml PGD, and set the loss to the trades's loss. But I cannot understand why it is and I guess I prefer to use criterion_kl(F.softmax(model(x_adv)), F.softmax(model(x_natural))). Are there some nice people who can answer my question? I would appreciate it.

I have the same question. And do you know more about that why the author use "criterion_kl" and "F.log_softmax(model(x_adv))"?