Deep-Mutual-Learning

TensorFlow implementation of Deep Mutual Learning accepted by CVPR 2018.

Introduction

Deep mutual learning provides a simple but effective way to improve the generalisation ability of a network by training collaboratively with a cohort of other networks.

Requirements

1. TensorFlow 1.3.1
2. CUDA 8.0 and cuDNN 6.0
3. Matlab

Usage

Data Preparation

1. Please download the Market-1501 Dataset

2. Convert the image data into TFRecords

sh scripts/format_and_convert_market.sh

Training

1. Train MobileNets with DML

sh scripts/train_dml_mobilenet_on_market.sh

2. Train MobileNet independently

sh scripts/train_ind_mobilenet_on_market.sh

Testing

1. Extract features of the test image

sh scripts/evaludate_dml_mobilenet_on_market.sh

2. Evaluate the performance with matlab code

Citation

If you find DML useful in your research, please kindly cite our paper:

@inproceedings{ying2018DML,
    author = {Ying Zhang and Tao Xiang and Timothy M. Hospedales and Huchuan Lu},
    title = {Deep Mutual Learning},
    booktitle = {CVPR},
    year = {2018}}