SHACIRA: Scalable HAsh-grid Compression for Implicit Neural Representations (ICCV 2023)

Official code for our paper
SHACIRA: Scalable HAsh-grid Compression for Implicit Neural Representations
Sharath Girish, Abhinav Shrivastava, Kamal Gupta
University of Maryland, College Park
ICCV 2023

This repository is built on top of an earlier version of Kaolin Wisp, a Pytorch library for working with neural fields. For a full detailed overview of running the scripts and their functions and capabilities, we recommend checking out their repository.

Setup

Clone repository

git clone git@github.com:Sharath-girish/shacira.git
cd shacira

Setup environment on conda

conda create -n "env_shacira" python=3.9.5
conda activate env_shacira

or virtualenv

python -m venv env_shacira
source env_shacira/bin/activate

Install Pytorch packages along with Kaolin Wisp packages

pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113
pip install kaolin==0.13.0 -f https://nvidia-kaolin.s3.us-east-2.amazonaws.com/torch-1.12.1_cu113.html

Build and install other dependencies as well as CUDA kernels for grid interpolation

pip install -r requirements.txt
pip install -r requirements_app.txt
python setup.py develop

Usage

The repository contains scripts for training INRs for images and NeRFs under the app folder with their respective config files containing the default argparse arguments.

Images

Script for training INRs on images. Input is a directory containing set of images to train different INRs on independently. Example run for the Kodak dataset.

python3 app/image/main_image.py --config app/image/configs/kodak.yaml --dataset-path /path/to/kodak/directory

NeRFs

Example script for training compressible feature grids on a scene in the RTMV dataset available at the dataset project page.

python3 app/nerf/main_nerf.py --config app/nerf/configs/nerf_latents.yaml --multiview-dataset-format rtmv --mip 2 --bg-color white --raymarch-type voxel --num-steps 16 --num-rays-sampled-per-img 4096 --dataset-num-workers 4 --dataset-path /path/to/scene

For the lego scene from the original NeRF dataset.

python3 app/nerf/main_nerf.py --config app/nerf/configs/nerf_latents.yaml --dataset-path /path/to/lego