lix-byte / DIHN

source code for paper "Deep Incremental Hashing Network for Efficient Image Retrieval" on CVPR-2019

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Deep Incremental Hashing Network for Efficient Image Retrieval

REQUIREMENTS

  1. pytorch>=1.0
  2. loguru

DATASETS

  1. CIFAR-10 Password: v1tj

  2. NUS-WIDE Password: uhr3

NUS-WIDE-Split-txt Password: xrqd

USAGE

usage: run.py [-h] [--dataset DATASET] [--root ROOT] [--batch-size BATCH_SIZE]
              [--lr LR] [--code-length CODE_LENGTH] [--max-iter MAX_ITER]
              [--max-epoch MAX_EPOCH] [--num-seen NUM_SEEN]
              [--num-samples NUM_SAMPLES] [--num-workers NUM_WORKERS]
              [--topk TOPK] [--gpu GPU] [--gamma GAMMA] [--mu MU]

DIHN_PyTorch

optional arguments:
  -h, --help            show this help message and exit
  --dataset DATASET     Dataset name.
  --root ROOT           Path of dataset
  --batch-size BATCH_SIZE
                        Batch size.(default: 64)
  --lr LR               Learning rate.(default: 1e-4)
  --code-length CODE_LENGTH
                        Binary hash code length.(default: 12)
  --max-iter MAX_ITER   Number of iterations.(default: 50)
  --max-epoch MAX_EPOCH
                        Number of epochs.(default: 3)
  --num-seen NUM_SEEN   Number of unseen classes.(default: 7)
  --num-samples NUM_SAMPLES
                        Number of sampling data points.(default: 2000)
  --num-workers NUM_WORKERS
                        Number of loading data threads.(default: 0)
  --topk TOPK           Calculate map of top k.(default: all)
  --gpu GPU             Using gpu.(default: False)
  --gamma GAMMA         Hyper-parameter.(default: 200)
  --mu MU               Hyper-parameter.(default: 50)

EXPERIMENTS

cifar-10: 7 original classes, 3 incremental classes.

nus-wide: 11 original classes, 10 incremental classes.

12 bits 24 bits 32 bits 48 bits
ADSH cifar-10 MAP@ALL 0.6433 0.6434 0.6451 0.6424
+DIHN cifar-10 MAP@ALL 0.9091 0.9117 0.9177 0.9217
ADSH nus-wide-tc21 MAP@5000 0.8056 0.8551 0.8594 0.8625
+DIHN nus-wide-tc21 MAP@5000 0.8361 0.9065 0.9028 0.9116

About

source code for paper "Deep Incremental Hashing Network for Efficient Image Retrieval" on CVPR-2019


Languages

Language:Python 100.0%