Generated labels

[dimasquest]

Is there a simple method to get newly generated labels (e.g. using the least_likely_attack) NOT from the model prediction, but rather from the atk method itself (for adversarial training)?

e.g. current labels are extracted using model(adv_images)

adv_images = atk(images, labels)
outputs = model(adv_images)

desired:

adv_images, adv_labels = atk(images, labels)

[Harry24k]

Thank you for your suggestion. Returning the predictions of adversarial labels is obviously useful, but it consumes one more forward for obtaining predictions of adversarial images. Thus, instead of atk.forward(), I added it to atk.save() as an argument save_pred. https://github.com/Harry24k/adversarial-attacks-pytorch/blob/6dbe9155b0ba6ff966f2d484366c13fcbf80e38d/torchattacks/attack.py#L149