Loss on codebook training. What are good loss values? about vqfr HOT 5 CLOSED

For 2): It depends on the task you would like to do. Sorry for not providing more useful information. If there are any other questions, feel free to reopen it.

from vqfr.

1. The loss seems to have some problems. Have you changed any configurations and do you run the code in batch=16? The following is our log at 28.5K.
   [epoch: 6, iter: 28,500, lr:(1.000e-04,)] [eta: 3 days, 12:00:50, time (data): 0.330 (0.003)] l_rec: 9.6236e-02 l_g_percep: 3.1971e-01 l_codebook: 6.3252e-02 l_total_g: 4.7919e-01
2. We decide the 800K following the configuration of GFP-GAN and Restoreformer. Since the reconstruction results will be stable after about 500K, maybe you can try smaller epochs. In 8 x A100 machines, the total codebook training cost is about 3 day, 13 hours. If you use V100 machines, it may requires about twice training time.

from vqfr.

I changed the dataset from FFHQ to full images instead of just faces and I changed the color space to LAB. I also changed the configuration as well as I have 4 GPUs not 8 and their max memory is 24,5 GB each. So the batch size per gpu is 3. Is it a must that it is in RGB color space? and can it manage full images instead of just faces?

from vqfr.

It seems you try to adapt this codebase to other tasks. Because we have not try other tasks, we cannot give informative instructions. But I can provide several hints for you: 1) the lpips perceptual loss receives input of RGB color space. 2) The batch size need to be large for real images, otherwise, it requires a smaller learning rate and enlarges training epochs (refer to VQGAN).

from vqfr.

1. Thank you so much, I can't believe I missed that.
2. To increase the batch size I must decrease the input dimensions, how much will this impact the quality do you think?

from vqfr.