Please explain trick 2

[mjdietzx]

https://github.com/soumith/ganhacks#2-a-modified-loss-function

When training generator standard way is to pass [batch_generated_imgs, batch_real_imgs] and np.ones(shape=batch_size * 2) where all images are labelled 1 (real) to trick discriminator.

If I understand this trick correctly it is saying to pass [batch_generated_imgs, batch_real_imgs] and np.concat(np.ones(shape=batch_size), np.zeroes(shape=batch_size)) where labels are now flipped for fake and real?

[spurra]

Trick 2 states that when training the generator, instead of minimising log(1-D(G(z))) you maximise log(D(G(z)) so that you get better gradients. This is because the discriminator usually performs better than the generator. This is most easily done when, for example in torch when using nn.CrossEntropyLoss, by assigning the synthetic samples the label 1 and training on that.

[soumith]

@mjdietzx you are correct in your understanding

[hellojialee]

Hi @mjdietzx @soumith, could you explain more about this. Should we always flip the label when training the Generator (if I understand correctly, the Discriminator is fixed at the same time). Thus, trick 2 [batch_real_imgs, np.zeroes(shape=batch_size] is actually destroying the Discriminator?