Potential bug of GAN example

[dolaameng]

Hi,

First thank you for sharing the tutorials!

I have a quick question on the GAN example, where 'detaching' the variable should be used in my opinion like fake_images = G(z).detach(), to avoid the training of Discriminator influences the Generator's behaviour at this stage. Or you have specific considerations here?

Thank you!

[yunjey]

@dolaameng No, it doesn't matter because we only train the discriminator using d_optimizer.step() and reset the grad buffer before calling g_loss.backward(). In the case of fake_images =Variable(G(z).detach().cuda()), it saves memory and speed up the training, but in my experience this is trivial.

[dolaameng]

Hi @yunjey yeah it makes sense. Thank you for clarification.