search

BadourAlBahar / pose-with-style

[SIGGRAPH Asia 2021] Pose with Style: Detail-Preserving Pose-Guided Image Synthesis with Conditional StyleGAN
https://pose-with-style.github.io/
MIT License
268 stars 41 forks source link

Discriminator loss too low #23

Closed tasinislam21 closed 2 years ago

tasinislam21 commented 2 years ago

I am using this model to train on my dataset. After a few epochs, the loss value of the discriminator gets low as 0.001, and the generator would be around 5-11. Is this normal?

BadourAlBahar commented 2 years ago

Please note that when the discriminator loss goes to zero this means that the discriminator is perfect at telling real from fake images. Check the real/fake images.. can you easily tell which one is real and which one is fake?

An example of when I had this issue happen is when the real image had zero side padding to make it square. The generated fake image can never generate exact zeros, best case ~0.0001. Hence, the discriminator can use this to easily distinguish real from fake images. To fix this, I also added zero padding to the fake image such that at the sides of both images would have exact zeros.

tasinislam21 commented 2 years ago

Thanks

Please note that when the discriminator loss goes to zero this means that the discriminator is perfect at telling real from fake images. Check the real/fake images.. can you easily tell which one is real and which one is fake?

An example of when I had this issue happen is when the real image had zero side padding to make it square. The generated fake image can never generate exact zeros, best case ~0.0001. Hence, the discriminator can use this to easily distinguish real from fake images. To fix this, I also added zero padding to the fake image such that at the sides of both images would have exact zeros.

Thanks for your response. There was a problem at my end. I did not remove the background of the real and fake images. Now I fixed that, and the error rate has been stabilised.