Notebooks

Three notebooks in notebooks/ run one signal end to end, one per task. They are the interactive version of the training scripts and the fastest way to see how the models compare on your own data.

Each compares FUTON against five representative baselines of the task at equal size: RFF, SIREN, FINER, Instant-NGP and TensoRF.

Notebook	Signal	Models	Runtime on an A100
image_representation.ipynb	data/Kodak/kodim19.png	7, about 197k parameters each, 1000 steps on the whole image	about 6 minutes
occupancy_volume.ipynb	data/occupancy/lucy.ply at \(256^3\)	7, about 131k parameters each, 1000 steps of \(2^{18}\) voxels	about 4 minutes
nerf.ipynb	data/nerf/blender/lego under FINER's protocol, in a quick pass	7, about 75k parameters each, 5000 steps	about 11 minutes (set NUM_STEPS = 25_000 for the benchmark's full schedule, about 35 minutes)

Run them

git clone https://github.com/pashtari/inro.git
cd inro
pip install -e ".[3d,nerf]" jupyterlab
python scripts/download_kodak.py --images 19
python scripts/download_meshes.py --meshes lucy
python scripts/download_blender.py --scenes lego
jupyter lab notebooks/

The notebooks read data from ../data/, relative to their folder, and are committed with their outputs, so they read without being run; rerun them top to bottom. The NeRF notebook also renders an orbit of its best INR as a GIF with imageio, which the nerf extra provides.

What is inside

Each notebook is self-contained and has the same sections:

Section	Content
Setup	Imports the packages, sets the seaborn theme and each model's color and marker, and defines the plotting helpers
Data	Shows the signal and the region the comparison magnifies
Models	Every model's constructor arguments and learning rate, with # deviation comments where a setting departs from the authors' code
Training	Runs inro.train or inro.nerf.train
Convergence	Plots the metric against steps and against training time
Summary	Tabulates every model's size, costs and scores, the best of each column in bold and the second underlined, and plots the scores against the costs
Qualitative Comparison	Sets the reconstructions side by side, magnified, above their errors

FUTON is drawn in black, and the baselines in seaborn's colorblind palette. The model settings are those of configs/<task>.yaml, which holds every model of the benchmark, the notebooks' and the rest.