search

Jittor / JNeRF

JNeRF is a NeRF benchmark based on Jittor. JNeRF re-implemented instant-ngp and achieved same performance with original paper.
Apache License 2.0
640 stars 74 forks source link

FFMLP(fp-16) not compatable with Jittor=13.8.5 #73

Open RobotDogCyberslacking opened 1 year ago

RobotDogCyberslacking commented 1 year ago

following the instructions in ReadMe.md , I got this when running the ngp_fox test:

/hy-tmp/JNeRF-master/python/jnerf/runner/runner.py:193: RuntimeWarning: invalid value encountered in cast
  ndarr = (img*255+0.5).clip(0, 255).astype('uint8')

The output image in log folder is all black. Some loss value is NAN.

my hardware environment: Intel(R) Xeon(R) CPU E5-2686 v4 RTX3090-24G

The training and output is all correct when I set using FP_16 = False in config file, which make the code use the MLP from pytorch.nn instead of FMLP.

I tried to fix it by downgrading my python to 3.8, jittor=1.3.4.13(exactly the version in requirements.txt),without running python setup.py, and everything goes fine.

If you encounter the same problem, try:

  1. downgrade your python to 3.8, jittor=1.3.4.13
  2. do not run python setup.py again, because it will upgrade your jittor package to the latest version automaticly.

I hope you guys fix it ASAP.

pechpo commented 1 year ago

I also encountered "RuntimeWarning: invalid value encountered in cast ndarr = (img*255+0.5).clip(0, 255).astype('uint8')" Using FP_16 = False solve my problem. Thank you for sharing.

scymz2 commented 5 months ago

jittor=1.3.4.13 doesn't support -arch=compute_89 (40-series GPU), need to mannually update cuda >=11.8, but I don't know how to do that.