asymmnet's Introduction

AsymmNet

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

AsymmNet: Towards ultralight convolution neural networks using asymmetrical bottlenecks

Video presentation on YouTube:

https://youtu.be/CMmZw8D15no

What's New

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

Block design

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

First, prepare data recordfile for ImageNet.
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

@article{yang2021asymmnet,
  title={AsymmNet: Towards ultralight convolution neural networks using asymmetrical bottlenecks},
  author={Yang, Haojin and Shen, Zhen and Zhao, Yucheng},
  journal={arXiv preprint arXiv:2104.07770},
  year={2021}
}