ashawkey / nerf2mesh

[ICCV2023] Delicate Textured Mesh Recovery from NeRF via Adaptive Surface Refinement

Home Page:https://me.kiui.moe/nerf2mesh/

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

sdf mode PSNR problem

trebladev opened this issue · comments

Hi,
I just use the scripts/runall_syn_sdf.sh to run sdf mode.
But I found that the PSNR of sdf mode is just 18 in chair scene, I get 34 without sdf mode.
Is that normal?
Thanks.

And when I train the ship, it will error below when epoch 40-45
loss=nan (nan) lr=0.007100: : 86% 86/100 [00:00<00:00, 98.30it/s]NaN or Inf found in input tensor. loss=nan (nan) lr=0.007100: : 87% 87/100 [00:00<00:00, 98.30it/s]NaN or Inf found in input tensor. loss=nan (nan) lr=0.007099: : 88% 88/100 [00:00<00:00, 98.30it/s]NaN or Inf found in input tensor. loss=nan (nan) lr=0.007099: : 89% 89/100 [00:00<00:00, 98.30it/s]NaN or Inf found in input tensor. loss=nan (nan) lr=0.007098: : 90% 90/100 [00:00<00:00, 98.30it/s]NaN or Inf found in input tensor. loss=nan (nan) lr=0.007098: : 91% 91/100 [00:00<00:00, 98.30it/s]NaN or Inf found in input tensor. loss=nan (nan) lr=0.007097: : 93% 93/100 [00:00<00:00, 98.93it/s]NaN or Inf found in input tensor. loss=nan (nan) lr=0.007097: : 93% 93/100 [00:00<00:00, 98.93it/s]NaN or Inf found in input tensor. loss=nan (nan) lr=0.007096: : 94% 94/100 [00:00<00:00, 98.93it/s]NaN or Inf found in input tensor. loss=nan (nan) lr=0.007096: : 95% 95/100 [00:00<00:00, 98.93it/s]NaN or Inf found in input tensor. loss=nan (nan) lr=0.007095: : 96% 96/100 [00:00<00:00, 98.93it/s]NaN or Inf found in input tensor. loss=nan (nan) lr=0.007094: : 97% 97/100 [00:01<00:00, 98.93it/s]NaN or Inf found in input tensor. loss=nan (nan) lr=0.007094: : 98% 98/100 [00:01<00:00, 98.93it/s]NaN or Inf found in input tensor.

commented

貌似sdf模式不仅降低了psnr的值,而且训练速度也很慢。

I also encountered the same problem NaN or Inf found in input tensor for certain examples. @trebladev Have you found the fix?

Same error here, at least with the synthetic scenes

I also encountered the same problem NaN or Inf found in input tensor for certain examples. @trebladev Have you found the fix?

same problem. Wondering reason. also be seen in #96

here is some result in lego
ngp_stage0_ep0060_0076_rgb_proc