iteratively generate adversarial samples
for n in range(args.T_adv):
input_aug_feat, output_aug = model.functional(params, False, input_aug, return_feat=True)
recon_batch_aug, _, = wae(input_aug)
# Constraint
constraint = mse_loss(input_feat, input_aug_feat)
ce_loss = criterion(output_aug, target_aug)
# Relaxation
relaxation = mse_loss(recon_batch, recon_batch_aug)
adv_loss = -(args.beta * relaxation + ce_loss - args.gamma * constraint)
aug_optimizer.zero_grad()
adv_loss.backward()
aug_optimizer.step()
The code of generate adversarial samples，which only trains wae and the value of input_aug has not changed. How can generate adversarial samples?
Trace the input_aug、virtual_test_images、only_virtual_test_images、X_aug、aug_dataset in the code, and no changes are found, but adversarial samples are formed, which is incomprehensible.
iteratively generate adversarial samples
for n in range(args.T_adv):
input_aug_feat, output_aug = model.functional(params, False, input_aug, return_feat=True)
recon_batch_aug, _, = wae(input_aug)
# Constraint
constraint = mse_loss(input_feat, input_aug_feat)
ce_loss = criterion(output_aug, target_aug)
# Relaxation
relaxation = mse_loss(recon_batch, recon_batch_aug)
adv_loss = -(args.beta * relaxation + ce_loss - args.gamma * constraint)
aug_optimizer.zero_grad()
adv_loss.backward()
aug_optimizer.step()
The code of generate adversarial samples，which only trains wae and the value of input_aug has not changed. How can generate adversarial samples?
Trace the input_aug、virtual_test_images、only_virtual_test_images、X_aug、aug_dataset in the code, and no changes are found, but adversarial samples are formed, which is incomprehensible.