Results reproduction about ngp_pl HOT 10 CLOSED

I think there is not a platform related issue here after I fixed vren compile error.

from ngp_pl.

Hi, are you using the latest code? Please set --scale 0.5. I'm still debugging for scale>0.5 (real scenes) and having difficulties, but all the nsvf data should work with default settings.

from ngp_pl.

Yep! I tried 0,5 also but got almost the same result. Is it a problem with optimiser?

from ngp_pl.

Can you first confirm that it works with the default optimizer? If so, then it's the problem of the optimizer. I've tried SGD, RAdam too and they didn't work either

from ngp_pl.

@kwea123 Thanks! Ill try apex optimizer, I did not do it yet because It is not easy to install it with Windows. I suppose that there is a problem with a regular Adam when we use mixed precision.

from ngp_pl.

Hi @kwea123! I've tried training wi9th the default optimizer and rusult does not change

from ngp_pl.

So you're using windows? I'll try when I have time, but bare in mind that this repo is for research purpose so if I can't find an easy way to solve this problem with windows, this issue will be "won't fix". Thanks for your comprehension.

from ngp_pl.

Sure! No problem!

from ngp_pl.

@kwea123 Thanks! Ill try apex optimizer, I did not do it yet because It is not easy to install it with Windows. I suppose that there is a problem with a regular Adam when we use mixed precision.

i replace apex FusedAdam to torch.optim.Adam and get psnr=34.30 on lego scene.

(base) D:\ngp_pl>python train.py --root_dir E:\Synthetic_NeRF\Synthetic_NeRF\Lego --exp_name Lego --scale 1 --batch_size 4096
GridEncoding: Nmin=16 b=1.38191 F=2 T=2^19 L=16
Using 16bit native Automatic Mixed Precision (AMP)
GPU available: True, used: True
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs
HPU available: False, using: 0 HPUs
GridEncoding: Nmin=16 b=1.38191 F=2 T=2^19 L=16
Using 16bit native Automatic Mixed Precision (AMP)
GPU available: True, used: True
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs
HPU available: False, using: 0 HPUs
Missing logger folder: logs\Lego                                                                                                                                                 508, train/s_per_ray=73.60, train/psnr=23.30]
Loading 100 train images ...
100it [00:03, 28.70it/s]
Loading 200 test images ...
200it [00:06, 30.06it/s]
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
Epoch 29: 100%|████████████████████████████████████████████████████| 1200/1200 
[01:47<00:00, 11.13it/s, loss=0.000213, train/s_per_ray=23.10, train/psnr=36.50, test/psnr=34.30]

from ngp_pl.

Thanks, btw you should use scale=0.5 for maximal performance on synthetic scenes

from ngp_pl.