README.md

1. Running example:

Environment: python 3

Requirements:

pytorch
torchvision

2. Statement:

As pytorch doesn't provide pretrained VGG-F model, unlike original DPSH paper, we use pretrained Alexnet or pretrained VGG-11 for feature learning part instead of pretrained VGG-F.

3. Data processing:

Following DPSH MatConvNet source code, we can obtain cifar-10.mat. To prepare data for pytorch version DPSH, run script ./data/CIFAR-10/SaveFig.m to save image files.

6. Demo:

python DPSH_CIFAR_10_demo.py

5. Result:

Mean Average Precision on CIFAR-10.

Net Structure	PlatForm	Code Length
		12 bits	24 bits	32 bits	48 bits
VGG-F	MatConvNet	0.713	0.727	0.744	0.757
Alexnet	Pytorch	0.7505	0.7724	0.7758	0.7828
VGG-11	Pytorch	0.7655	0.8042	0.8070	0.8108

Training Loss on CIFAR-10.

6. Influence of Hyper-Parameter \lambda