GraphBit: Bitwise Interaction Mining via Deep Reinforcement Learning This repo contains the basic code for Graphbit on CIFAR-10.

The code is based on Keras and TensorFlow. Please prepare the real-valued feature in 4096 dimensions of CIFAR-10 by downing the data through the link, and run main.py to train the network and get hashed compact features. Finally you can use bifeat_extract.m to extract the binary feature and run retrival.m to get the mAP on retrival task.

The repo is organized as follows:

• Python: the code written by python
  • main.py: The main
  • US_network.py: Contains Unsupervised Network and its necessary loss function.
  • RL_network.py: Deep Q Network, its actions, rewards and state matrix.
• Matlab: the test file
  • bifeat_extract.m: extract binary feature from real-valued feature.
  • precision.m: calculate the precision.
  • retrival.m: calculate the mAP of retrival task based on binary feature.
• TXT: two files are the train and test label. Need to mention that the order of features are not the same as in https://www.cs.toronto.edu/~kriz/cifar.html.
• Link for downloading input VGG feature:
  • dropbox:
    • train: https://www.dropbox.com/s/f930w9vga62nv1i/feat16_train.npy?dl=0
    • test: https://www.dropbox.com/s/lpa7jeuel4jdo80/feat16_test.npy?dl=0
  • baidu:
    • train: https://pan.baidu.com/s/1v_Q7P388YoJ4yFjycuZ29w
    • test: https://pan.baidu.com/s/1Dlk07xqxH9nQOgqgUTbkEA

If you find Graphbit useful, please cite it.

@inproceedings{duan2018graphbit,
title={GraphBit: Bitwise Interaction Mining via Deep Reinforcement Learning},
author={Duan, Yueqi and Wang, Ziwei and Lu, Jiwen and Lin, Xudong and Zhou, Jie},
booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
pages={8270--8279},
year={2018}
}