jjh13 / ncdl Goto Github PK

NCDL is a software package for differentiable computation on non-Cartesian lattices. But what is a "non-Cartesian lattice"? Have you ever wondered why pixels are square? Why not use Hexagons instead; the neighbor structure of a hexagon is more uniform than a square. Notably, using a hexagonal structure has a higher capacity for information compared to a square structure [1,2,3]. There are similar structures in higher dimensions that allow

This package provides a means of performing computation on these structures. We implement these operations in PyTorch, and can therefore provide differentiable computation for no additional development cost.

It is worth mentioning that this is a research codebase. As such, our focus is heavily weighted towards correctness.

This is the inital release of NCDL. The core functionality of the lattice tensor, plus all of the operations mentioned in the paper are implemented. Most network architecture structures that are defined on Cartesian data can be relatively easily moved to a non-Cartesian sapce.

There have been a few things cut out of this version that have still not been merged back into the repo:

• Max/atrous pooling is missing its cuda implementation.

Changes are currently happening in master, but we'll eventually switch to a more sane development practice.

First, make sure PyTorch is installed. There's no specific requirement for this in the setup.py file; this is because NCDL is relatively agnostic to the version of PyTorch used.

Next, clone the repository via

git clone https://github.com/jjh13/NCDL.git

Finally, in the root of this directory, install to the local environment with

pip install  -e .

Check our readthedocs page, https://ncdl.ai. Currently, the key functionality is documented, but there may be gaps. Please open an issue if you find any documentation lacking. There are relatively comprehensive examples in the examples directory. Please take a look at those.

Documentation is in docs/build/index.html. It's worth looking at for some clarity on what's currently available in NCDL.

See modules/autoencoder.py. It has a number of models defined. Some are successful experiments that didn't make it into the current version of the paper, some are less successful. That's a full worked neural network implemented with NCDL.

Experiments are in the experiments folder. You will need pytorch lighting 1.8.1 to run them. The configs are in the configs folder (you'll also need the.

If you use this work in your publication, be sure to cite

@inproceedings{horacsek2023ncdl,
  title={NCDL: A Framework for Deep Learning on non-Cartesian Lattices},
  author={Horacsek, Joshua John and Alim, Usman},
  booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
  year={2023}
}