Could you explain a little bit about the statistic metrics? about basicsr HOT 5 CLOSED

Hi @Jing25
D_fake: The mean of D outputs for fake ones (generated SR images)
D_real: The mean of D outputs for real ones (GT)
l_d_fake: D loss for fake ones (generated SR images)
l_d_real: D loss for real ones (GT)
l_g_fea: G loss (feature loss, e.g., percetual loss)
l_g_gan: G loss (adversarial loss )
l_g_pix: G loss (pixel loss)
Run the codes, and get their values.

You may also need to change
"which_model_G": "RRDB_net" // RRDB_net | sr_resnet
in the test_esrgan.json.

from basicsr.

How can I test SRGAN with your code? In the test_esrgan.json file, I changed "model" into "srgan" and "pretrain_model_G" into "SRGAN_bicx4_303_505.pth". But there's error.

from basicsr.

If D_fake/D_real is the mean of D outputs fake/real, then I guess the expected optimal value of each of them should be 0.5?
I think G loss (l_g_fea, l_g_gan and l_g_pix) should decrease with number of iteration increasing. How about D loss (l_d_fake and l_d_real)?

Yeah. I change "which_model_G" to "sr_resnet" and then "nb" to 16. It works. But I'm just curious why I cannot change "norm_type": null to "batch"? SRGAN uses batch normalization. But why here if I change "norm_type" to "batch" it has error?

from basicsr.

1. In vanilla GAN, the mean of D outputs measures the realness or fakeness of the inputs. Thus, in theory, the D_fake and D_real should converge to both 0.5 (See https://arxiv.org/pdf/1406.2661.pdf ). However, in practice, this is not the case. For example, you can find the D outputs for DCGAN in https://github.com/carpedm20/DCGAN-tensorflow, where D_real approaches to 1 and D_fake approaches to 0.
   Furthermore, the outputs of D are not necessarily the scores of realness or fakeness.
   In ESRGAN, we use Relativistic GAN, where the discriminator predict
   relative realness instead of the absolute value. The meaning of D outputs is not clear.

I attached one example for ESRGAN, D_real and D_fake are the values before sigmoid.

2. The SRGAN model we provided is different from the official one. We also omit the BN layers and do some slight changes. But the main ideas -- using ResNet stucture, vgg+adv loss, are the same.

from basicsr.

Thank you for the explanation. It helps a lot.

from basicsr.