NeuralCompression

What's New

• August 2023 (image compression) - Released PyTorch implementation of MS-ILLM
• April 2023 (video compression) - Released PyTorch implementation of VCT
• November 2022 (image compression) - Released Bits-Back coding with diffusion models!

About

NeuralCompression is a Python repository dedicated to research of neural networks that compress data. The repository includes tools such as JAX-based entropy coders, image compression models, video compression models, and metrics for image and video evaluation.

NeuralCompression is alpha software. The project is under active development. The API will change as we make releases, potentially breaking backwards compatibility.

Installation

NeuralCompression is a project currently under development. You can install the repository in development mode.

PyPI Installation

First, install PyTorch according to the directions from the PyTorch website. Then, you should be able to run

pip install neuralcompression

to get the latest version from PyPI.

Development Installation

First, clone the repository and navigate to the NeuralCompression root directory and install the package in development mode by running:

pip install --editable ".[tests]"

If you are not interested in matching the test environment, then you can just apply pip install -e ..

Repository Structure

We use a 2-tier repository structure. The neuralcompression package contains a core set of tools for doing neural compression research. Code committed to the core package requires stricter linting, high code quality, and rigorous review. The projects folder contains code for reproducing papers and training baselines. Code in this folder is not linted aggressively, we don't enforce type annotations, and it's okay to omit unit tests.

The 2-tier structure enables rapid iteration and reproduction via code in projects that is built on a backbone of high-quality code in neuralcompression.

neuralcompression

• neuralcompression - base package
  • data - PyTorch data loaders for various data sets
  • distributions - extensions of probability models for compression
  • functional - methods for image warping, information cost, flop counting, etc.
  • layers - building blocks for compression models
  • metrics - torchmetrics classes for assessing model performance
  • models - complete compression models
  • optim - useful optimization utilities

projects

• projects - recipes and code for reproducing papers
  • bits_back_diffusion - code for bits-back coding with diffusion models
  • deep_video_compression DVC (Lu et al., 2019), deprecated
  • illm MS-ILLM (Muckley et al., 2023)
  • jax_entropy_coders - implementations of arithmetic coding and ANS in JAX
  • torch_vct VCT (Mentzer, et al.,)

Tutorial Notebooks

This repository also features interactive notebooks detailing different parts of the package, which can be found in the tutorials directory. Existing tutorials are:

• Walkthrough of the neuralcompression flop counter (view on Colab).
• Using neuralcompression.metrics and torchmetrics to calculate rate-distortion curves (view on Colab).

Contributions

Please read our CONTRIBUTING guide and our CODE_OF_CONDUCT prior to submitting a pull request.

We test all pull requests. We rely on this for reviews, so please make sure any new code is tested. Tests for neuralcompression go in the tests folder in the root of the repository. Tests for individual projects go in those projects' own tests folder.

We use black for formatting, isort for import sorting, flake8 for linting, and mypy for type checking.

License

NeuralCompression is MIT licensed, as found in the LICENSE file.

Model weights released with NeuralCompression are CC-BY-NC 4.0 licensed, as found in the WEIGHTS_LICENSE file.

Some of the code may from other repositories and include other licenses. Please read all code files carefully for details.

Cite

If you use code for a paper reimplementation. If you would like to also cite the repository, you can use:

@misc{muckley2021neuralcompression,
    author={Matthew Muckley and Jordan Juravsky and Daniel Severo and Mannat Singh and Quentin Duval and Karen Ullrich},
    title={NeuralCompression},
    howpublished={\url{https://github.com/facebookresearch/NeuralCompression}},
    year={2021}
}