readme.md

GhostNet

GhostNet: More Features from Cheap Operations. CVPR 2020. [arXiv]

By Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, Chang Xu.

• Approach

• Performance

GhostNet beats other SOTA lightweight CNNs such as MobileNetV3 and FBNet.

Implementation

This repo provides the TensorFlow code and pretrained model of GhostNet on ImageNet. The PyTorch implementation can be found at https://github.com/iamhankai/ghostnet.pytorch.

myconv2d.py implemented GhostModule and ghost_net.py implemented GhostNet.

Requirements

The code was verified on Python3.6, TensorFlow-1.13.1, Tensorpack-0.9.7. Not sure on other version.

Usage

Run python main.py --eval --data_dir=/path/to/imagenet/dir/ --load=./models/ghostnet_checkpoint to evaluate on val set.

You'll get the accuracy: top-1 error=0.26066, top-5 error=0.08614 with only 141M Flops (or say MAdds).

Data Preparation

ImageNet data dir should have the following structure, and val and caffe_ilsvrc12 subdirs are essential:

dir/
  train/
    ...
  val/
    n01440764/
      ILSVRC2012_val_00000293.JPEG
      ...
    ...
  caffe_ilsvrc12/
    ...

caffe_ilsvrc12 data can be downloaded from http://dl.caffe.berkeleyvision.org/caffe_ilsvrc12.tar.gz

Citation

@inproceedings{ghostnet,
  title={GhostNet: More Features from Cheap Operations},
  author={Han, Kai and Wang, Yunhe and Tian, Qi and Guo, Jianyuan and Xu, Chunjing and Xu, Chang},
  booktitle={CVPR},
  year={2020}
}

Other versions

This repo provides the TensorFlow code of GhostNet. Other versions can be found in the following:

0. Pytorch: code
1. Darknet: cfg file, and description
2. Gluon/Keras/Chainer: code
3. Pytorch for human pose estimation: code