iie-mr/fmdh

source code for paper "Fast and Multilevel Semantic-Preserving Discrete Hashing" on BMVC-2019

A pytorch implementation for paper "Fast and Multilevel Semantic-Preserving Discrete Hashing" BMVC-2019

REQUIREMENTS

1. python >= 3.5.2
2. pytorch >= 0.4

DATASETS

1. MIRFLICKR-25K
2. MSCOCO: train val

USAGE

usage: fmdh_mirflickr.py [-h] [--bits BITS] [--gpu GPU] [--arch ARCH]
                         [--max-iter MAX_ITER] [--epochs EPOCHS]
                         [--batch-size BATCH_SIZE] [--topk TOPK] [--m M]
                         [--m_wap M_WAP] [--C C] [--num-samples NUM_SAMPLES]
                         [--alpha ALPHA] [--beta BETA]
                         [--learning-rate LEARNING_RATE]

fmdh_mirflickr

optional arguments:
  -h, --help            show this help message and exit
  --bits BITS           binary code length (default: 16,32,64)
  --gpu GPU             selected gpu (default: 0)
  --arch ARCH           model name (default: resnet152)
  --max-iter MAX_ITER   maximum iteration (default: 20)
  --epochs EPOCHS       number of epochs (default: 1)
  --batch-size BATCH_SIZE
                        batch size (default: 64)
  --topk TOPK           top k (default: 100)
  --m M                 ndcg@m (default: 100)
  --m_wap M_WAP         ndcg@m (default: 100)
  --C C                 class number (default: 24)
  --num-samples NUM_SAMPLES
                        hyper-parameter: number of samples (default: 2000)
  --alpha ALPHA         hyper-parameter: alpha (default: 100000)
  --beta BETA           hyper-parameter: beta (default: 100)
  --learning-rate LEARNING_RATE
                        hyper-parameter: learning rate (default: 0.001)

RESULT

NDCG@100

Dataset	Code Length
	16 bits	32 bits	64 bits
MIRFLICKR-25K	0.5517	0.5793	0.6028
MSCOCO	0.5036	0.5370	0.5541

ACG@100

Dataset	Code Length
	16 bits	32 bits	64 bits
MIRFLICKR-25K	2.501	2.739	2.760
MSCOCO	1.501	1.618	1.633

Weighted MAP@100

Dataset	Code Length
	16 bits	32 bits	64 bits
MIRFLICKR-25K	2.536	2.723	2.722
MSCOCO	1.530	1.697	1.708

This code originates from https://github.com/jiangqy/ADSH-AAAI2018/tree/master/ADSH_pytorch