Official PyTorch implementation of paper HHF: Hashing-guided Hinge Function for Deep Hashing Retrieval
NVIDIA GPU + CUDA (may CuDNN) and corresponding PyTorch framework
Python 3
We recommend you to follow https://github.com/thuml/HashNet/tree/master/pytorch#datasets to prepare ImageNet and COCO images.
Please run the training command with --dataset cifar10/cifar100 directly and the cifar10/cifar100 dataset will be downloaded automatically.
python retrieval.py --dataset [dataset] --backbone [backbone] --method [method] --hash_bit [hash_bit]
Arguments (default value)
--dataset: dataset [(imagenet), cifar10, cifar100, coco]
--backbone: backbone network for feature extracting [(googlenet), resnet]
--method: baseline method [(anchor), NCA, DHN]
--hash_bit: length of hash bits [(48), or any positive integer that ≤ 256]
Other optional arguments (default value)
--alpha: a hyper-parameter to control the gradient of the metric loss [(16), or any positive float value]
--beta: a hyper-parameter to balance the contribution of metric loss and quantization loss [(0.001), or any positive float value]
--delta: a relaxation hyper-parameter to alleviate the overfitting problem [(0.2), or any positive float value that < 1]
--batch_size the size of a mini-batch [(85), or any positive integer]
Add --test after the training command. Make sure there is a corresponding .ckpt file in the ./result/ directory.
python retrieval.py --dataset [dataset] --backbone [backbone] --method [method] --hash_bit [hash_bit] --test
If you use this method or this code in your research, please cite as:
@misc{xu2022hhf,
title={HHF: Hashing-guided Hinge Function for Deep Hashing Retrieval},
author={Chengyin Xu and Zenghao Chai and Zhengzhuo Xu and Hongjia Li and Qiruyi Zuo and Lingyu Yang and Chun Yuan},
year={2022},
eprint={2112.02225},
archivePrefix={arXiv},
primaryClass={cs.CV}
}