Skip to content

[CVPRW 2021] AsymmNet: Towards ultralight convolution neural networks using asymmetrical bottlenecks.

License

Notifications You must be signed in to change notification settings

yanghaojin/AsymmNet

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

10 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

AsymmNet

This repo is the official implementation with MXNet/Gluon for:

Video presentation on YouTube:

What's New

  • 2021.08.18 release MXNet trained weight && log.txt && mac.txt
  • 2021.07.06 pytorch version released

Block design

Structure compare

Requirements

  • Python >= 3.6
  • MXNet >= 1.5.1
  • GluonCV >= 0.7.0
  • mxboard >= 0.1.0
  • scipy
  • tqdm
  • portalocker

What is included

├── README.md
├── models
│   ├── asymmnet.py
│   ├── common.py
│   ├── ghostnet.py
│   ├── mobilenetv2.py
│   └── mobilenetv3.py
├── utils
│   ├── args_helper.py
│   ├── utils.py
│   └── visualization.py
└── train.py

How to use

Referred gluoncv_examples_classification

For training ImageNet

  1. First, prepare data recordfile for ImageNet.
  2. Training
export MXNET_SAFE_ACCUMULATION=1 && export MXNET_ENABLE_GPU_P2P=1
python train.py \
    --dataset imagenet \
    --rec-train /path/to/train.rec \
    --rec-train-idx /path/to/train.idx \
    --rec-val /path/to/val.rec \
    --rec-val-idx /path/to/val.idx \
    --batch-size 256 --model asymmnet_large \
    --mode hybrid --wd 0.00003 \
    --lr 2.6 --lr-mode cosine --num-epochs 360 --num-gpus 8 -j 48 \
    --warmup-epochs 5  --no-wd --last-gamma \
    --label-smoothing --width-scale 1.0 \
    --dtype float16 --dropout 0.1 \
    --tag-name 1.0x-fp16-8gpu-lr2.6-360e-drop0.1 \
    --log-interval 100

Citation

@InProceedings{Yang_2021_CVPR,
    author    = {Yang, Haojin and Shen, Zhen and Zhao, Yucheng},
    title     = {AsymmNet: Towards Ultralight Convolution Neural Networks Using Asymmetrical Bottlenecks},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2021},
    pages     = {2339-2348}
}

Reference

About

[CVPRW 2021] AsymmNet: Towards ultralight convolution neural networks using asymmetrical bottlenecks.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%