caffe-moon: Face attributes predict with Mixed Objective Optimization

Introduction

This repo provide a solution for face attributes prediction, I design a tiny neural network with Mixed Objective Optimizition, And trained on celebA face database with mis-aligned face. I get 89+% accuracy, better than paper's result. The trained model is only 4+M size.

network structure

This network structure is inspired by yolo-tiny

Usage

Train

The script for prepare the sample will update after I clean up my code.
1, Put this repo in "caffe-root/examples/"
2, Modify the path in "solver" , "train_val.prototxt" and "train.sh"
3, cd "caffe-root/"
4, sh ./examples/caffe-moon/train.sh"

Test

I provide some simple python code for test. You should install ipython befor run "test_moon_euclidean.ipynb"

Coming soon...

1, finetuning with 5 attributes
2, face landmark with face attributes
3, ...

Reference

1, MOON: A Mixed Objective Optimization Network for the Recognition of Facial Attributes