Scalable Low-Rank Compression for Neural Networks

This repository contains the Tensorflow implementation of our paper. The implementation can reproduce the result on CIFAR-10.

If you find our work useful in your research, please consider citing:

@inproceedings{ijcai2021-447,
  title     = {Decomposable-Net: Scalable Low-Rank Compression for Neural Networks},
  author    = {Yaguchi, Atsushi and Suzuki, Taiji and Nitta, Shuhei and Sakata, Yukinobu and Tanizawa, Akiyuki},
  booktitle = {Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, {IJCAI-21}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},
  editor    = {Zhi-Hua Zhou},
  pages     = {3249--3256},
  year      = {2021},
  month     = {8},
  note      = {Main Track}
  doi       = {10.24963/ijcai.2021/447},
  url       = {https://doi.org/10.24963/ijcai.2021/447},
}

Environment

• 1 GPU with more than 8GB RAM (recommended)
• Ubuntu 18.04 LTS
• Anaconda 4.7.12 (or newer)

Usage

1. Create and activate conda env:

conda create -n tf1-py3 tensorflow-gpu=1.15.0 anaconda
conda activate tf1-py3

2. Prepare the dataset:

sh prepare_cifar10.sh

3. Run training:

sh run_cifar10.sh {gpu_id} ./dataset/cifar10/train.txt ./dataset/cifar10/test.txt vgg15 0.01 0.25 KyKxCin_Cout 0