Transmute AI Model Efficiency Toolkit

Introduction

Trailmet is a model efficiency toolkit for compressing deep learning models using state of the art compression techniques. Today deep learning models are not deployable because of their huge memory footprint, TRAILMET is an effort to make deep learning models more efficient in their size to performance ratio. It is developed using Pytorch 1.13.

Major features

• State of the art compression algorithms implemented.
• Demo notebooks for training each algorithm.
• Modular Design: All alogithms are modular and can customized easily for any kind of model and dataset.

Installation

Below are quick steps for installation:

git clone https://github.com/transmuteAI/trailmet.git
cd trailmet
conda create -n trailmet
conda activate trailmet
conda install pytorch=1.13 torchvision=0.14 pytorch-cuda=11.7 -c pytorch -c nvidia
pip install trailmet

Algorithms Implemented

Demo notebooks for each algorithm is added in experiments folder

Knowledge Distillation

• Response KD
• Factor Transfer
• Attention Transfer

Pruning

• Chipnet
• Network slimming
• Growth Regularization

Quantization

• BitSplit
• BRECQ
• LAPQ

Binarization

• BiRealNet
• ReActNet
• BNN-BN

Acknowledgement

Citation

If you find this project useful in your research, please consider cite:

@misc{,
    title={},
    author={},
    howpublished = {}},
    year={2023}
}

License

This project is released under the MIT license.