Hashing Models

This repository includes the adversarial autoencoder models for the following papers:

• Image Hashing by Minimizing Discrete Component-wise Wasserstein Distance (arxiv'20)
• Efficient Implicit Unsupervised Text Hashing using Adversarial Autoencoder (WWW'20)
• Adversarial Factorization Autoencoder for Look-alike Modeling (CIKM'19)

Please cite the appropriate paper(s) if you are using this repository. If you have any questions, please also feel free to send us an email at khoadoan@vt.edu.

What does this reposistory have?

This repository includes the following features:

• Code to create the common experimental evaluations for the learning-to-hash task,
• Various implemented ranking metrics for the learning-to-hash task.
• Visualization code.
• Several neural net models that were used in the original papers. Note: due to time constraint, we will add/update the models gradually
• Adversarial training algorithm for learning the hash functions.

1. Setup

Prerequisites

We need the following:

• conda or miniconda (preferred)
• GPU or CPU
• your perseverance 😃

Install the environment

Clone the repository and run the following commands inside the basedir of this repository:

conda create --name hashing --file requirements.txt python=3.7
conda activate hashing

and you are ready to have some fun. Your perseverance stops here! Wait, more like "your perseverance starts here" 😁!!!

To run the jupyer notebook

You can start the jupyter notebook server by invoking the following command:

CUDA_VISIBLE_DEVICES=<gpu ids> bin/run_notebook <id>

This will open a notebook server at port 888 on your machine.

If you don't have a GPU, you can run the notebook on the CPU by the following command.

bin/run_notebook_cpu

This will open a notebook server at port 8888 on your machine if it's available.

Repository artifacts

• python: all the python code will be here
• notebooks: all the notebooks will be here
• bin: all the executables will be here
• data: data will be here (but it's not checked in)
• code: all the external code references will be here (e.g. external dependent repos)
• requirements.txt: list of python reqs
• README.md: this doc, and light documentation of this repos.

2. Create the experiment data

The notebook notebooks/create_datasets.ipynb contains the following tasks:

• Randomly split the datasets into train, db, query sets. There are different sizes of the splits, each of which is specific to some papers in the literature.
• The creations of the following datasets do not need any manual downloads: mnist, fashion-mnist, cifar10. These datasets are downloaded automatically using torchvision.
• The creation of the following datasets need manual downloads into the data folder before we can split them: flickr25k, place365.

Note: for flickr25k, download the dataset at this link.

To extract the VGG-pretrained features, use the notebook extract_VGG_features.ipynb. We can extract the features at different intermediate layers of the VGG network.

3. Run Aversarial Autoencoderss

The experiments on each dataset will be in the notebook notebooks/ae-<dataset>.ipynb.