/lightweight-neural-architecture-search

This is a collection of our zero-cost NAS and efficient vision applications.

Primary LanguagePythonApache License 2.0Apache-2.0

News

Introduction

English | 简体中文

Lightweight Neural Architecture Search (Light-NAS) is an open source zero-short NAS toolbox for backbone search based on PyTorch. The master branch works with PyTorch 1.4+ and OpenMPI 4.0+.

Major features

License

This project is developed by Alibaba and licensed under the Apache 2.0 license.

This product contains third-party components under other open source licenses.

See the NOTICE file for more information.

Changelog

1.0.3 was released in 2022/11/30:

  • Support MAE-NAS for DAMO-YOLO.
  • Add 2 new blocks: k1kx and kxkx.
  • Add a tutorial on NAS for DAMO-YOLO.

Please refer to changelog.md for details and release history.

Installation

Prerequisites

  • Linux
  • GCC 7+
  • OpenMPI 4.0+
  • Python 3.6+
  • PyTorch 1.4+
  • CUDA 10.0+

Prepare environment

  1. Compile the OpenMPI 4.0+ Downloads.

    cd path
tar -xzvf openmpi-4.0.1.tar.gz
cd openmpi-4.0.1
./configure --prefix=/your_path/openmpi
make && make install

    add the commands into your ~/.bashrc

    export PATH=/your_path/openmpi/bin:$PATH
export LD_LIBRARYPATH=/your_path/openmpi/lib:$LD_LIBRARY_PATH

  2. Create a conda virtual environment and activate it.

    conda create -n light-nas python=3.6 -y
conda activate light-nas

  3. Install torch and torchvision with the following command or offcial instruction.

    pip install torch==1.4.0+cu100 torchvision==0.5.0+cu100 -f https://download.pytorch.org/whl/torch_stable.html

    if meet "Not be found for jpeg", please install the libjpeg for the system.

    sudo yum install libjpeg # for centos
sudo apt install libjpeg-dev # for ubuntu

  4. Install other packages with the following command.

    pip install -r requirements.txt

Easy to use

  • Search with examples

    cd scripts/classification
sh example_xxxx.sh

Results

Results for Classification, Details are here.

Backbone size Param (M) FLOPs (G) Top-1 Structure Download
R18-like 224 10.8 1.7 78.44 txt model
R50-like 224 21.3 3.6 80.04 txt model
R152-like 224 53.5 10.5 81.59 txt model

Note: If you find this useful, please support us by citing it.

@inproceedings{zennas,
	title     = {Zen-NAS: A Zero-Shot NAS for High-Performance Deep Image Recognition},
	author    = {Ming Lin and Pichao Wang and Zhenhong Sun and Hesen Chen and Xiuyu Sun and Qi Qian and Hao Li and Rong Jin},
	booktitle = {2021 IEEE/CVF International Conference on Computer Vision},  
	year      = {2021},
}

The official code for Zen-NAS was originally released at https://github.com/idstcv/ZenNAS.

Results for Object Detection, Details are here.

Backbone Param (M) FLOPs (G) box APval box APS box APM box APL Structure Download
ResNet-50 23.5 83.6 44.7 29.1 48.1 56.6 - -
ResNet-101 42.4 159.5 46.3 29.9 50.1 58.7 - -
MAE-DET-S 21.2 48.7 45.1 27.9 49.1 58.0 txt model
MAE-DET-M 25.8 89.9 46.9 30.1 50.9 59.9 txt model
MAE-DET-L 43.9 152.9 47.8 30.3 51.9 61.1 txt model

Note: If you find this useful, please support us by citing it.

@inproceedings{maedet,
  title     = {MAE-DET: Revisiting Maximum Entropy Principle in Zero-Shot NAS for Efficient Object Detection},
  author    = {Zhenhong Sun and Ming Lin and Xiuyu Sun and Zhiyu Tan and Hao Li and Rong Jin},
  booktitle = {International Conference on Machine Learning},
  year      = {2022},
}

Results for low-precision backbones, Details are here.

Backbone Param (MB) BitOps (G) ImageNet TOP1 Structure Download
MBV2-8bit 3.4 19.2 71.90% - -
MBV2-4bit 2.3 7 68.90% - -
Mixed19d2G 3.2 18.8 74.80% txt model
Mixed7d0G 2.2 6.9 70.80% txt model

Note: If you find this useful, please support us by citing it.

@inproceedings{qescore,
  title     = {Entropy-Driven Mixed-Precision Quantization for Deep Network Design on IoT Devices},
  author    = {Zhenhong Sun and Ce Ge and Junyan Wang and Ming Lin and Hesen Chen and Hao Li and Xiuyu Sun},
  journal   = {Advances in Neural Information Processing Systems},
  year      = {2022},
}

Main Contributors

Zhenhong Sun, Ming Lin, Xiuyu Sun, Hesen Chen, Ce Ge, Yilun Huang.