Skip to content

ccetaw/tensorf_pl

Repository files navigation

TensorRF Pytorch Lightning

This is a fork of TensorRF, reimplemented in Pytorch Lightning. With this implementation, we achieve beter modulization and faster training.

Create a conda environment using environment.yml. It's tested on cuda 11.7 and corresponding torch version.

To train

python launch.py --config ./configs/lego.txt --train

To run testing

python launch.py --config /path/to/saved/config.txt --test --resume /path/to/ckpt

Structure

The original structure is decomposed to:

  • Density field
  • Radiance field
  • Occupancy grid
  • A renderer that implements the rendering algorithm

This decomposition makes modules reusable and could be further used for other applications such as inverse rendering, semantic segmentation.

Results

G.T. Original This repo

For lego scene, on a single RTX 2080, with batch size 4096 and 30000 iters

time psnr
Original 25:30 35.51
PL 22:14 35.53

Other metrics(LPIPS, SSIM) are similar to values reported in original paper.

About

TensoRF implemented in Pytorch Lightning

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages