mah-pytorch's Introduction

MAH-Pytorch

PyTorch implementation of our paper "Collaborative Learning for Extremely Low Bit Asymmetric Hashing" [Link].

Preparation

Dependencies

Python 2.7
PyTorch (version >= 0.4.1)

Datasets

CIFAR download the CIFAR-10 Matlab version [Link] then run the script

matlab ./data/CIFAR-10/SaveFig.m

NUSWIDE dataset
MIRFlickr dataset

(referenced repo)

Results

Mean Average Precision on CIFAR-10.

Method	Backbone	Code Length
Method	Backbone	4 bits	6 bits	8 bits	10 bits	12 bits
MAH-flat	ResNet50	0.4759	0.8197	0.9339	0.9335	0.9503
MAH-cascade	ResNet50	0.7460	0.8950	0.9429	0.9489	0.9537

Precision@5K Figures

Usage

For traning with the cascaded multihead structure on different datasets:

python cascade_CIFAR-10.py --bits '4' --gpu '1' --batch-size 64
python cascade_FLICKR.py --bits '4' --gpu '1' --batch-size 64
python cascade_NUS_WIDE.py --bits '4' --gpu '1' --batch-size 64

For traning with the flat multihead structure on different datasets:

python flat_CIFAR-10.py --bits '4' --gpu '1' --batch-size 64
python flat_FLICKR.py --bits '4' --gpu '1' --batch-size 64
python flat_NUS_WIDE.py --bits '4' --gpu '1' --batch-size 64

Citation

Please cite the following paper in your publications if it helps your research:

@article{DBLP:journals/corr/abs-1809-09329,
    author    = {Yadan Luo and
                 Yang Li and
                 Fumin Shen and
                 Yang Yang and
                 Peng Cui and
                 Zi Huang},
    title     = {Collaborative Learning for Extremely Low Bit Asymmetric Hashing},
    journal   = {CoRR},
    volume    = {abs/1809.09329},
    year      = {2018},
    url       = {http://arxiv.org/abs/1809.09329},
    archivePrefix = {arXiv},
    eprint    = {1809.09329},
    timestamp = {Wed, 13 Mar 2019 15:40:02 +0100},
    biburl    = {https://dblp.org/rec/bib/journals/corr/abs-1809-09329},
    bibsource = {dblp computer science bibliography, https://dblp.org}
  }

Recommend Projects