karlszp / cghash

A pytorch implementation of our paper: "Unsupervised Hashing with Contrastive Learning by Exploiting Similarity Knowledge and Hidden Structure of Data" [ACM MM 2023]

CGHash: Conditional Generative Model enhanced Contrastive Hashing

A pytorch implementation of our paper:

Unsupervised Hashing with Contrastive Learning by Exploiting Similarity Knowledge and Hidden Structure of Data

Accepted in Proceedings of the 31st ACM International Conference on Multimedia [ACM MM 2023].

Training Schema

Settings

1. Place downloaded datasets at datasets/. (Cifar-10 will be downloaded automatically.)
2. [Optional] Refer to utils/data_path.py to set dataset path.
3. Place pre-trained models at models/pretrained_backbones/. (A 300-epoch SimCLR pre-trained models for Cifar-10 is available here.)
4. Configure training details in configs/[dataset]/[stage]_[dataset]_[code_length].yml.

Training Example for 64-bit Cifar-10

1. Place Cifar-10 dataset at datasets/ manually or download automatically later.

2. Configure training details in configs/cifar-10/mine_cifar10.yml, configs/cifar-10/cghash_cifar10_64.yml, configs/cifar-10/selflabel_cifar10_64.yml.

3. Run shell training script.

   chmod +x ./run.sh
   ./run.sh [CODE_LENGTH] [GPU_ID]
   # e.g. ./run.sh 64 0,1

Retrieval Example