ShuffleNet V1&V2

this code is mxnet implementation of ShuffleNetV1 and ShuffleNetv2, For details, please read the original paper:
shufflenetv1
shufflenetv2
This code is based on farmingyard's implementation(https://github.com/farmingyard/ShuffleNet)

Code is test on MxNet 1.11.0

Installation

Clone this repository, and we'll call the directory that you cloned mxnet-shufflenet as ${SHUFFLENET_ROOT}.

git clone https://github.com/Tveek/mxnet-shufflenet.git

Install shuffle channel operator to MXNet:

2.1 Clone MXNet and checkout to MXNet by

git clone --recursive https://github.com/dmlc/mxnet.git
git submodule update

2.2 Copy operators in $(SHUFFLENET_ROOT)/source/shuffle_channel*.xxx by

cp -r $(SHUFFLENET_ROOT)/operator/* $(MXNET_ROOT)/src/operator/contrib/

2.3 Compile MXNet

cd ${MXNET_ROOT}
make -j $(nproc) USE_OPENCV=1 USE_BLAS=openblas USE_CUDA=1 USE_CUDA_PATH=/usr/local/cuda USE_CUDNN=1

2.4 Install the MXNet Python binding by

cd python
sudo python setup.py install

Python operator function are also in symbol file, so you can use it without above

RGB mean and std are used(rgb_mean=[123.68,116.779,103.939], rgb_std=[58.393,57.12,57.375])
The top-1/5 accuracy rates by using single random crop (crop size: 224x224, image size: 256xN)

Network	Top-1	Top-5	model size
ShuffleNet_V1_1x3	63.94	85.27	7.1MB
ShuffleNet_V2_1	65.43	86.50	8.7MB