treezzz / dsdh_pytorch

Source code for paper "Deep Supervised Discrete Hashing" on NeurIPS-2017

Deep Supervised Discrete Hashing

REQUIREMENTS

pip install -r requirements.txt

1. pytorch
2. loguru

DATASETS

1. CIFAR-10
2. NUS-WIDE Password: uhr3
3. Imagenet100 Password: ynwf

USAGE

usage: run.py [-h] [--dataset DATASET] [--root ROOT] [--batch-size BATCH_SIZE]
              [--arch ARCH] [--lr LR] [--code-length CODE_LENGTH]
              [--max-iter MAX_ITER] [--num-query NUM_QUERY]
              [--num-train NUM_TRAIN] [--num-workers NUM_WORKERS]
              [--topk TOPK] [--gpu GPU] [--mu MU] [--nu NU] [--eta ETA]
              [--evaluate-interval EVALUATE_INTERVAL]

DSDH_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: 128)
  --arch ARCH           CNN model name.(default: alexnet)
  --lr LR               Learning rate.(default: 1e-5)
  --code-length CODE_LENGTH
                        Binary hash code length.(default: 12,24,32,48)
  --max-iter MAX_ITER   Number of iterations.(default: 150)
  --num-query NUM_QUERY
                        Number of query data points.(default: 1000)
  --num-train NUM_TRAIN
                        Number of training data points.(default: 5000)
  --num-workers NUM_WORKERS
                        Number of loading data threads.(default: 6)
  --topk TOPK           Calculate map of top k.(default: all)
  --gpu GPU             Using gpu.(default: False)
  --mu MU               Hyper-parameter.(default: 1e-2)
  --nu NU               Hyper-parameter.(default: 1)
  --eta ETA             Hyper-parameter.(default: 1e-2)
  --evaluate-interval EVALUATE_INTERVAL
                        Evaluation interval.(default: 10)

EXPERIMENTS

CNN model: Alexnet. Compute mean average precision(MAP).

cifar10: 1000 query images, 5000 training images.

nus-wide-tc21: 21 classes, 2100 query images, 10500 training images.

Imagenet-tc100: 100 classes, 5000 query images, 10000 training images. mu=5, nu=5, eta=2.

bits	12	16	24	32	48	64	128
cifar10@ALL	0.6994	0.7052	0.7204	0.7267	0.7345	0.7466	0.7493
nus-wide-tc21@5000	0.8052	0.8155	0.8317	0.8332	0.8437	0.8478	0.8544
imagenet-tc100@1000	0.1622	0.1793	0.2356	0.2808	0.3350	0.3753	0.4274

dsdh_pytorch's Issues

关于nuswide数据集的小问题~

打扰一下，新手对于数据集的使用还不大懂~ 请问nuswide数据集下载了之后是里面5个4g的文件都要用到吗？还是只用到了那个1.86G的zip，解压用这个就好呢？您分享的nus-wide-21是论文里常用的，只包含21类的nus-wide的子数据集吗？

Do you have the trained models available?

Hi! I'd like to try the models as is on a dataset, do you have the pretrained Pytorch models available? I'd love to try it without having to train my own on ImageNet. Thanks!

RMSprop

树哥你好，我把网络换了resnet50，发现还真的只有RMSprop的结果能正常一点点，Adam和SGD好像没法正常更新。Any insight on this？仅是讨论哈哈，不是Issue。

Are you sure you really implemented the paper?

I wana know which paper you implemented

multi gpus training

Environment: pytorch==1.1.0 torchvision==0.3.0 cudatoolkit=9.0
Problem:
I added two lines:
model = nn.DataParalle(model)
model.to(device) after model = AlexNet(code_length) in alexnet.py, but I found it can not speed up the training, can you help me?

NUSWIDE数据集问题

请问下NUSWIDE数据集下，这两行的写法是不是有误？正确的写法应该是：

query_index = perm_index[:num_query]
train_index = perm_index[num_query:num_query+num_train]

DSDH_PyTorch/data/nus_wide.py

Lines 70 to 71 in 4c1f5a0

	query_index = perm_index[:num_query]
	train_index = perm_index[:num_train]