evaluate D about esrgan

[xinntao]

I think that D cannot make absolutely correct predictions as D is also improving itself during training.
Typical losses of G and D can be found here: https://github.com/carpedm20/DCGAN-tensorflow (see Training details part)

I usually print the D outputs to see whether D is trained OK. If the outputs change suddenly or drastically, D usually behaviors badly.

[JimmyChame]

Hi, I have another doubt about GAN. @xinntao
When I use GAN and perceptual loss to finetune my network, which is pre-trained with mse loss, I save the outputs of the generator during training. Then I observe that the outputs become textured, then smooth, then textured, and so on. Is this a normal phenomnon during training ? Is the best result achieved at the end of training ? Thank you !

[xinntao]

@JimmyChame
At first of training, it will fluctuate a lot. With training, the outputs will be more stable.

[JimmyChame]

Thanks for your reply. I have another question if it doesn't interrupt. @xinntao
I check the losses you mentioned in DCGAN-tensorflow, but I have a little doubt about them.
In the ideal case, the discriminator cannot distinguish between real and generated images at the end of training. The outputs of discriminator will be around 0.5, which means both of real and fake logits are 0. So I calculate the losses myself as following:

real = tf.ones_like(input)*0
fake = tf.ones_like(input)*0
real_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=real, labels=tf.ones_like(real)))
fake_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=fake, labels=tf.zeros_like(fake)))
d_loss = real_loss+fake_loss
g_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=fake, labels=tf.ones_like(fake)))

Then I get the d_loss and g_loss values which are 1.3855876 and 0.6927938 separately. I found a large deviation between the value in DCGAN-tensorflow and the one I calculated. Am I wrong, or is this just an ideal case ?

[xinntao]

@JimmyChame It is just the ideal case.