Skip to content

Installation

Python 3.11 or newer and PyTorch 2.10 or newer, with or without CUDA. Everything runs on the CPU; a GPU makes training practical.

Install

pip install inro

pip install "inro[3d,nerf]"

For the training scripts, configs and notebooks, which live in the repository:

git clone https://github.com/pashtari/inro.git
cd inro
pip install -e ".[3d,nerf,dev]"

Check the install:

import torch, inro

print(inro.__version__, torch.cuda.is_available())

Extras

Extra Installs Needed for
3d PyMCubes, pycollada, trimesh, Open3D Meshes as occupancy volumes, and exporting and rendering reconstructed meshes
nerf nerfacc, imageio CUDA occupancy-grid ray marching, and writing novel-view videos
dev black, pytest, pyinstrument Formatting, the test suite and profiling
docs mkdocs-material, mkdocs-glightbox, mkdocstrings Building this site

Two versions that matter

  • trimesh is pinned to 5.1.1, which voxelized the benchmark's meshes: other versions move the grids, and IoU by up to 0.2 points.
  • nerfacc compiles a CUDA extension the first time it is used and needs a CUDA toolkit for that. Without one, inro.nerf.is_nerfacc_available() is False and inro.nerf.create_renderer() falls back to the built-in PyTorch renderer, which produces the same images more slowly.

Benchmark data

The experiments use three public datasets. The scripts, in the repository, download them into data/, which is ignored by git.

Dataset Script Content Size Subset
Kodak scripts/download_kodak.py 24 images, 768 × 512 about 30 MB --images 1 17 23
Stanford scripts/download_meshes.py 5 meshes about 3 GB --meshes armadillo thai_statue
Blender scripts/download_blender.py 8 NeRF synthetic scenes about 2.4 GB --scenes lego ship

Each script also takes --output-dir and --force. Interrupted downloads resume, and existing files are kept. The layout is what the training scripts and notebooks expect:

data/
├── Kodak/kodim01.png ... kodim24.png
├── occupancy/armadillo.ply  dragon.ply  happy_buddha.ply  lucy.ply  thai_statue.ply
└── nerf/blender/<scene>/transforms_{train,val,test}.json + images

Occupancy volumes are voxelized on first use and cached beside the mesh in data/occupancy/.cache/<name>_<resolution>.pt.

Development

From a clone of the repository:

pip install -e ".[dev]"
black --check inro tests scripts  # formatting, line length 88
pytest -q                         # tests that need CUDA or local data skip themselves