ESRGAN: incomplete loss_GAN

[lukas-blecher]

Hi, I think the loss_GAN in esrgan.py is missing a term. In the ESRGAN paper the adversarial loss is like the discriminator loss a sum of two terms. However as far as I understand the code, only the second part is implemented in esrgan.py#L135. The following line would do the trick in my opinion.

loss_GAN = (criterion_GAN(pred_fake - pred_real.mean(0, keepdim=True), valid) + 

            criterion_GAN(pred_real - pred_fake.mean(0, keepdim=True), fake))/2