A re-implementation of sphereface.

Introduction

The operation of user-defined op(marginInnerProduct) is completely the same as it in the sphereface-caffe.

So the operation of computing Backward gradient also uses the tricks sphereface used.

Requirements

Linux (I've tested on ubuntu16.04)
Tensorflow1.4+(?)
boost cuda(if you have installed caffe, there is no problem  for those libraries)
matlab

Installation

1. First:

git clone https://github.com/hujun100/tensorflow-sphereface.git 	
cd tensorflow-sphereface

2. Change the path variable TENSORFLOW and CUDA_DIR in "makefile" according to your configuration Then:

make

Train

1. Change the configuration in config.py.

cd train_test
python train.py

Test(LFW)

1. Put the aligned face images to "lfw_evaluation" folder
2. Change the test configuration in config.py
3. Change the test_data_dir in evaluation.m

python get_lfw_features.py
cd ../lfw_evaluation
run evaluation.m

Result(LFW)

1. I've get 98.10% accuracy on LFW trained on CASIA-webface. I believe a better result will be obtained if I use data augmentation(mirror, smooth, jpeg compression and so on)

Reference

sphereface

caffe

Thanks to all the contributors of sphereface and caffe  

Others 

Data augmentation

Now I provide the c++ code of some useful data augmentation(shift, zoom, rotation, modHSV, modRGB and so on) in data_augmentation.hpp. You can write python code according to it to improve performance.