Chainer SSDH

This repository contains the experimental implementation of Supervised Semantics-preserving Deep Hashing model using Chainer framework. This model can generate the semantics-preserving binary code from raw image, and it can be trained as simple classification task.

NOTE: This is not the official implementation.

Requirements

• Python

• Chainer

• Scipy

• Jupyter (For Demo)

• Matplotlib (For Demo)

Training

At first, please download the pre-trained model parameter of AlexNet. Download script is provided.

$ bash scripts/download_alexnet.sh

Then, convert the caffemodel file to npz format to save the initialization time of training script.

$ python scripts/convert_caffemodel_to_npz.py

Now, let's start training of SSDH model. Length of output binary code can be specified by --units or -u option.

$ python code/train.py -g 0 --out output/unit48 --units 48

Please check the all options by using --help option.

Demo

The sample notebook for similar image search is provided under the notebooks directory. Here is the sample output of image search using SSDH model.