This is a caffe repository for learning to hash. We fork the repository from Caffe and make our modifications. The main modifications are listed as follow:

• Add multi label layer which enable ImageDataLayer to process multi-label dataset.
• Add pairwise loss layer and quantization loss layer described in paper "Deep Hashing Network for Efficient Similarity Retrieval".

In data/nus_wide/train.txt, we give an example to show how to prepare training data. In data/nus_wide/parallel/, the list of testing and database images are splitted to 12 parts, which could be processed parallelly when predicting.

In models/DHN/nus_wide/, we give an example to show how to train hash model. In this model, we use pairwise loss and quantization loss as loss functions.

The bvlc_reference_caffenet is used as the pre-trained model. If the NUS_WIDE dataset and pre-trained caffemodel is prepared, the example can be run with the following command:

"./build/tools/caffe train -solver models/DHN/nus_wide/solver.prototxt -weights models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel"

In pairwise loss layer and quantization loss layer, parameter loss_weight can be tuned to give them different weights.

In models/DHN/predict/predict_parallel.py, we give an example to show how to evaluate the trained hash model.

@inproceedings{DBLP:conf/aaai/ZhuL0C16,
  author    = {Han Zhu and
               Mingsheng Long and
               Jianmin Wang and
               Yue Cao},
  title     = {Deep Hashing Network for Efficient Similarity Retrieval},
  booktitle = {Proceedings of the Thirtieth {AAAI} Conference on Artificial Intelligence,
               February 12-17, 2016, Phoenix, Arizona, {USA.}},
  pages     = {2415--2421},
  year      = {2016},
  crossref  = {DBLP:conf/aaai/2016},
  url       = {http://www.aaai.org/ocs/index.php/AAAI/AAAI16/paper/view/12039},
  timestamp = {Thu, 21 Apr 2016 19:28:00 +0200},
  biburl    = {http://dblp.uni-trier.de/rec/bib/conf/aaai/ZhuL0C16},
  bibsource = {dblp computer science bibliography, http://dblp.org}
}