search

olivia23333 / E3Gen

MIT License
52 stars 3 forks source link

Inconsistent results of my trained model and the official checkpoint #2

Open shenwenhao01 opened 2 months ago

shenwenhao01 commented 2 months ago

Thanks for your great work!

I followed the training guidance and trained the model on 2 GPUs for 160,000 iters. But the rendered evaluation results look quite noisy and different from your checkpoint.

Below are the evaluation results under the viz_uncond folder.

scene_0000_005 scene_0001_002 scene_0002_024 scene_0263_025

olivia23333 commented 2 months ago

Terribly sorry for that, I set bg_color=1 (indicates white background) in lines 47 and 59 of the configuration file ssdnerf_avatar_uncond_thuman_conv_16bit.py during training, and in the ICON part used in data preprocessing, the version I uploaded renders the scan in a black background, which causes the noisy training results you encounter. There are two ways to solve this problem:

  1. Change bg_color=1 to bg_color=0 in the configuration file. (There are two in the configuration file, both need to be changed) This way, gt and the rendered generated results have the same black background.
  2. Change glClearColor(0.0, 0.0, 0.0, 0.0) to glClearColor(1.0, 1.0, 1.0, 0.0) in ICON/lib/renderer/gl/render.py in the ICON section (lines 248 and 328), then re-render the scan to get the rendered ground truth with a white background and retrain the network.

Also, during training the PSNR in iteration 10000 should be close to 30.8. During retraining, if the PSNR in iteration 10000 shows much lower than this, you can update the results and I'll check if there are any other errors.

Thanks for the issue! I'll change ICON to the correct version tomorrow, as the server containing this code is under repair today.

shenwenhao01 commented 2 months ago

Thanks for your timely reply! I will try again.