This repository contains Condensa Lite, a stripped-down version of Condensa that supports pruning and learning-rate rewinding.

Status: Condensa Lite is under active development, and bug reports, pull requests, and other feedback are all highly appreciated. See the contributions section below for more details on how to contribute.

Condensa Lite requires:

• A working Linux installation (we use Ubuntu 20.04)
• NVIDIA drivers and CUDA 11+ for GPU support
• Python 3.6 or newer
• PyTorch 1.8 or newer

Retrieve the latest source code from the Condensa repository:

git clone --branch lite https://github.com/NVlabs/condensa.git

Navigate to the source code directory and run the following:

pip install -e .

To check the installation, run the unit test suite:

bash run_all_tests.sh -v

Documentation for Condensa Lite is available here. Please also check out the original Condensa paper for a detailed description of Condensa's motivation, features, and performance results.

We appreciate all contributions, including bug fixes, new features and documentation, and additional tutorials. You can initiate contributions via Github pull requests. When making code contributions, please follow the PEP 8 Python coding standard and provide unit tests for the new features. Finally, make sure to sign off your commits using the -s flag or adding Signed-off-By: Name<Email> in the commit message.

If you use Condensa Lite for research, please consider citing the following paper:

@article{condensa:2020,
    title={A Programmable Approach to Neural Network Compression},
    author={Joseph, Vinu and Gopalakrishnan, Ganesh L
            and Muralidharan, Saurav and Garland, Michael and Garg, Animesh},
    journal={IEEE Micro},
    volume={40},
    number={5},
    pages={17--25},
    year={2020},
    publisher={IEEE}
}

Condensa Lite is a research prototype and not an official NVIDIA product. Many features are still experimental and yet to be properly documented.