HiDDeN
HiDDeN copied to clipboard
training a question
local predGeneratorLoss = criterion_adv:forward(predFake, 2 * labels_encoded) local predGeneratorGradOut = criterion_adv:backward(predFake, 2 * labels_encoded) local predGeneratorGradIn = adversary:backward(payloads_encoded, predGeneratorGradOut, 0) gradOutput[1] = gradOutput[1] + opt.adversary_gradient_scale * predGeneratorGradIn
I don't know why to train the discriminator to treat the encoded picture as the real picture