Reproducing performance on DTU

[sxyu]

Reproducing performance on DTU

[danperazzo]

Hello @hongsukchoi, I think I managed to reproduce it, or at least got close to it. I trained in the default setting (only with the batch size reduced by 2 from the default 4) and took about 300~ epochs to converge (about 4-5 days with an RTX 2080Ti). My model got a PSNR of 18.499 and an SSIM of 0.679 on the 3 views DTU evaluation.

[colorful-liyu]

Hello @hongsukchoi, I think I managed to reproduce it, or at least got close to it. I trained in the default setting (only with the batch size reduced by 2 from the default 4) and took about 300~ epochs to converge (about 4-5 days with an RTX 2080Ti). My model got a PSNR of 18.499 and an SSIM of 0.679 on the 3 views DTU evaluation.

hello, I am trying to reproduce it on the DTU, too. But every epoch takes about 2.5 hours with 3 RTX 2080Ti, for it repeats DTU 32 times in one epoch. I was wondering how could it run 300 epochs during 4-5 days with an RTX 2080Ti.

[TianyuDai]

Hello @hongsukchoi, I think I managed to reproduce it, or at least got close to it. I trained in the default setting (only with the batch size reduced by 2 from the default 4) and took about 300~ epochs to converge (about 4-5 days with an RTX 2080Ti). My model got a PSNR of 18.499 and an SSIM of 0.679 on the 3 views DTU evaluation.

Hello @danperazzo , thank you for your information! Could I ask why do you reduce batch size from 4 to 2?

[eiyike123]

Hello @hongsukchoi, I think I managed to reproduce it, or at least got close to it. I trained in the default setting (only with the batch size reduced by 2 from the default 4) and took about 300~ epochs to converge (about 4-5 days with an RTX 2080Ti). My model got a PSNR of 18.499 and an SSIM of 0.679 on the 3 views DTU evaluation.

I am finding difficulty running pixel-nerf

[eiyike123]

This is the error i got trying to run pixel-nerf

pyparsing.exceptions.ParseSyntaxException: Expected '}', found '=' (at char 759), (line:34, col:18)

[GateraGael]

Hello all, this was resolved in the following thread #62

pip install --force-reinstall -v "pyhocon==0.3.55"

[MathewCrowley]

Hello @hongsukchoi, I think I managed to reproduce it, or at least got close to it. I trained in the default setting (only with the batch size reduced by 2 from the default 4) and took about 300~ epochs to converge (about 4-5 days with an RTX 2080Ti). My model got a PSNR of 18.499 and an SSIM of 0.679 on the 3 views DTU evaluation.

Hello! When I try to run eval.py, it cannot find con/resnet_fine_mv.conf. How did you solve this problem?