khoadoan/single_loss_quantization

Implementations of Deep Hashing Models

Prerequisites

We need the following:

• conda or miniconda (preferred)
• GPU or CPU

Setup the environment

Clone the repository. The setup script to initialize and activate the environment is collected in etc/setup_env. Simply run the following command:

. etc/setup_env

Datasets

The NUS-WIDE and COCO datasets can be downloaded here. The CIFAR10 dataset can be downloaded automatically with PyTorch.

Repository artifacts

• python: code folder
• requirements.txt: list of python reqs
• README.md: this doc, and light documentation of this repos.

Using Original Deep Hashing Methods

• CIFAR10: nohup python python/HashNet.py --dataset cifar10 --data_root data/ --random_seed 9 --bit_list 16 32 64 128 --model AlexNet --epochs 200 --optimizer adam --lr 1e-5 --test_every 10 --save_path experiments/HashNet/cifar10_AlexNet_b64_adam 2>&1 >experiments/logs/HashNet-r9-cifar10_AlexNet_b64_adam.log &
• COCO: nohup python python/HashNet.py --dataset coco --data_root data/ --random_seed 9 --bit_list 16 32 64 128 --model AlexNet --epochs 200 --optimizer adam --lr 1e-5 --test_every 10 --save_path experiments/HashNet/coco_AlexNet_b64_adam 2>&1 >experiments/logs/HashNet-r9-coco_AlexNet_b64_adam.log &
• NUS-WIDE: nohup python python/HashNet.py --dataset coco --data_root data/ --random_seed 9 --bit_list 16 32 64 128 --model AlexNet --epochs 200 --optimizer adam --lr 1e-5 --test_every 10 --save_path experiments/HashNet/nuswide_21_AlexNet_b64_adam 2>&1 >experiments/logs/HashNet-r9-nuswide_21_AlexNet_b64_adam.log &

We currently support the following methods

• Cao et al. HashNet: Deep Learning to Hash by Continuation. ICCV 2017. [Paper] [HashNet.py]
• Li et al. Deep Supervised Discrete Hashing. NIPS 2017. [Paper] [DSDH.py]

Using Distributional Quantization Approach

Doan et al. One Loss for Quantization: Deep Hashing with Discrete Wasserstein Distributional Matching (CVPR2022). [Paper]

• This repository supports various quantization losses discussed in Doan et al. For HSWD, use --quantization_type swdC; for SWD, use --quantization_type swd; we also support Optimal Transport estimation using --quantization_type ot.

• CIFAR10: nohup python python/HashNet.py --dataset cifar10 --data_root data/ --random_seed 9 --bit_list 16 32 64 128 --model AlexNet --epochs 200 --optimizer adam --lr 1e-5 --quantization_type swdC --quantization_alpha 0.1 --test_every 10 --save_path experiments/HashNet/cifar10_AlexNet_b64_adam 2>&1 >experiments/logs/HashNet-r9-cifar10_AlexNet_b64_adam_swdC.log &

• COCO: nohup python python/HashNet.py --dataset coco --data_root data/ --random_seed 9 --bit_list 16 32 64 128 --model AlexNet --epochs 200 --optimizer adam --lr 1e-5 --quantization_type swdC --quantization_alpha 0.1 --test_every 10 --save_path experiments/HashNet/coco_AlexNet_b64_adam 2>&1 >experiments/logs/HashNet-r9-coco_AlexNet_b64_adam_swdC.log &

• NUS-WIDE: nohup python python/HashNet.py --dataset coco --data_root data/ --random_seed 9 --bit_list 16 32 64 128 --model AlexNet --epochs 200 --optimizer adam --lr 1e-5 --quantization_type swdC --quantization_alpha 0.1 --test_every 10 --save_path experiments/HashNet/nuswide_21_AlexNet_b64_adam 2>&1 >experiments/logs/HashNet-r9-nuswide_21_AlexNet_b64_adam_swdC.log &

Citations

Please cite the following work when using this repository:

@inproceedings{doan2022one,
  title={One Loss for Quantization: Deep Hashing with Discrete Wasserstein Distributional Matching},
  author={Doan, Khoa D and Yang, Peng and Li, Ping},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={9447--9457},
  year={2022}
}

Acknowledgements

• This respository is inspired by this respository https://github.com/swuxyj/DeepHash-pytorch. Thank you the authors of DeepHash-pytorch.