Deep Progressive Quantization

This is the python implementation of Beyond Product Quantization: Deep Progressive Quantization for Image Retrieval (IJCAI-19 accepted). Please refer to paper link for detailed infomation. The python scripts include following experiments:

• Training a model and Encoding a database main.py
• Evaluating AQD mAP on query set Eval.py
• Performing tSNE transform on CIFAR-10 tSNE.py
• Evaluating classification accuracy Accuracy.py
• Checking time consumption Time.py

Prerequisites

The scripts require following package:

• tensorflow (>= 1.4.1, either cpu or gpu version)
• numpy (>=1.14.3)
• scikit-learn (>=0.19.1)
• opencv-python (>=3.4)

The scripts have been tested on a machine equipped with:

• CPU: Intel(R) Xeon(R) CPU E5-2620 V4
• GPU: NVIDIA Titan X Pascal
• RAM: 128 GB (Please note that we didn't optimize memory usage, so the memory consumption would be very large)
• OS: Ubuntu 16.04 LTS

Please download AlexNet pretrianed model from ()()() and place it into ./data/models/

Dataset Preparation

CIFAR-10:

• Training set: 5000
• Query set: 1000
• Gallery: 54000

You should split CIFAR-10 by yourself. Download it and split it as the paper described.

---

NUS-WIDE:

• Training set: 10000
• Query set: 5000
• Gallery: 190834

Please download the dataset and put it into a specific directory. Then you should modify the prefixs of all paths in /data/nus21

---

Imagenet:

• Training set: 10000
• Query set: 5000
• Gallery: 128564

Please download the dataset (ILSVRC >= 2012) and put it into a specific directory. Then you should modify the prefixs of all paths in /data/imagenet

Run

Firstly, list all the parameter of scripts:

python main.py --help

usage: main.py [-h] [--Dataset DATASET] [--Mode MODE] [--BitLength BITLENGTH]
           [--ClassNum CLASSNUM] [--K K] [--PrintEvery PRINTEVERY]
           [--LearningRate LEARNINGRATE] [--Epoch EPOCH]
           [--BatchSize BATCHSIZE] [--Device DEVICE] [--UseGPU [USEGPU]]
           [--noUseGPU] [--SaveModel [SAVEMODEL]] [--noSaveModel] [--R R]
           [--Lambda LAMBDA] [--Tau TAU] [--Mu MU] [--Nu NU]

optional arguments:
-h, --help            show this help message and exit
--Dataset DATASET     The preferred dataset, 'CIFAR', 'NUS' or 'Imagenet'
--Mode MODE           'train' or 'eval'
--BitLength BITLENGTH
                        Binary code length
--ClassNum CLASSNUM   Label num of dataset
--K K                 The centroids number of a codebook
--PrintEvery PRINTEVERY
                        Print every ? iterations
--LearningRate LEARNINGRATE
                        Init learning rate
--Epoch EPOCH         Total epoches
--BatchSize BATCHSIZE
                        Batch size
--Device DEVICE       GPU device ID
--UseGPU [USEGPU]     Use /device:GPU or /cpu
--noUseGPU
--SaveModel [SAVEMODEL]
                        Save model at every epoch done
--noSaveModel
--R R                 mAP@R, -1 for all
--Lambda LAMBDA       Lambda, decribed in paper
--Tau TAU             Tau, decribed in paper
--Mu MU               Mu, decribed in paper
--Nu NU               Nu, decribed in paper

To perform a training:

python main.py --Dataset='CIFAR' --ClassNum=10 --LearningRate=0.001 --Device=0

To perform an evaluation:

python Eval.py --Dataset='CIFAR' --ClassNum=10 --Device=0 --R=-1

Citations

Please use the following bibtex to cite our papers:

@inproceedings{DPQ,
  title={Beyond Product Quantization: Deep Progressive Quantization for Image Retrieval},
  author={Gao, Lianli and Zhu, Xiaosu and Song, Jingkuan and Zhao, Zhou and Shen, Heng Tao},
  booktitle={Proceedings of the 2019 International Joint Conferences on Artifical Intelligence (IJCAI)},
  year={2019}
}

@inproceedings{DRQ,
  title={Deep Recurrent Quantization for Generating Sequential Binary Codes},
  author={Song, Jingkuan and Zhu, Xiaosu and Gao, Lianli and Xu, Xin-Shun and Liu, Wu and Shen, Heng Tao},
  booktitle={Proceedings of the 2019 International Joint Conferences on Artifical Intelligence (IJCAI)},
  year={2019}
}