NHL

The codes for a paper, we provides our codes and some deep hashing model baselines.

Main Dependencies

• pytohn 3.8
• torch 1.11.0+cu113
• numpy 1.22.4
• pandas 2.0.3

How to run

You can easily run our code by following steps:

• Replace "{your root}" in the file "utils/tools.py" with your own file path.
• We have prepered some cases in scripts/main.sh, you counld run the command "sh scripts/main.sh" to begin the training process.
• Note that MDSH models should fisrt build hash centers. We have prepred their hash centers in tmp_file/*, or you can use bit_length/MDSH_cg.py to build the hash centers of MDSH.

The explanations of main options

• --device: choose the used cuda.
• --dataset: select a dataset from [cifar10, imagenet, coco]
• --info: choose a deep hashing model [CSQ, DBDH, DCH, DHN, DPN, DSH, DTSH, LCDSH, SHCIR, MDSH]
• --mode: with or without NHL
• --analysis: if True, use adaptive weight strategy
• --distill: if True, use long-short cascade self-distillation
• --distill_weight: the weight for long-short cascade self-distillation