CompressAI – a new PyTorch library for deep learning compression research was released this week. The new library aims to provide researchers with a flexible tool for doing research and contributing to the deep learning compression domain.
Built on top of PyTorch, the library features custom operations, layers, and models for deep learning-based data compression, along with a model zoo with exported compression models and evaluation scripts for the models included in the library. The library comes with a set of examples, scripts, and notebooks for beginners to start learning its usage.
As part of CompressAI’s Model Zoo, researchers included a few ready-to-use models: mshj2018_factorized, bmshj2018_hyperprior, mbt2018_mean, mbt2018, cheng2020_anchor, and cheng2020_attn.
The library code was open-sourced and it is available on Github along with details on how to set it up and use it. More comprehensive documentation is available here with details about the library API, utils, and the model zoo.
CompressAI was developed by researchers and engineers from InterDigital AI Lab and it is licensed under the Apache 2.0 license.