HashGAN-Pytorch

Introduction

This is a pytorch implementation of Unsupervised Deep Generative Adversarial Hashing Network, CVPR'18 for CIFAR10 and MNIST dataset.

Prerequisites

• Linux

  This code was written to be run on Linux.

• Python > 3.6

  Using conda is recommended: https://docs.anaconda.com/anaconda/install/linux/

• pytorch

  To install: https://pytorch.org/get-started/locally/

• numpy

  Run this in terminal: conda install numpy

• tqdm

  Run this in terminal: conda install tqdm

Installation

Run this in terminal:

git clone https://github.com/8uos/HashGAN-Pytorch

Usage

Train

The simplest way is just run this in terminal:

python HashGAN.py --train

There are some possible additional arguments:

--batch_size: The size of minibatch. Default value is 100.

--z_dim : The length of a continuous part of input random variable. Default value is 128.

--b_dim : The length of a binary part of input random variable. Default value is 16.

--init_lr : Initial learning rate. Default value is 9e-04.

--final_lr : Final learning rate. Default value is 3e-04.

--epochs : The number of epochs to train. Default value is 100.

--gpu : The id of gpu to use. Default value is 0.

--dataset : The name of dataset to use. 'cifar10' and 'mnist' are possible, and default value is 'mnist'.

--G_dict : The path to generator model to load. If this is None, generator will be randomly initialized. Default value is None.

--D_dict : The path to discriminator model to load. If this is None, discriminator will be randomly initialized. Default value is None.

--save_dir : The name of the save directory. Everything will be saved in results/save_dir. Default value is 'temp'.

--log_step : Interval to print the losses.

--data_dir : The location of the dataset. If the dataset does not exists in data_dir, the dataset will be downloaded. Default value is './data'.

Evaluate

python HashGAN.py --eval \
                  --G_dict=Path/to/generator/dict/to/evaluate \
                  --D_dict=Path/to/discriminator/dict/to/evaluate

The possible additional arguments are identical to the ones above.

Structure

HashGAN.py

• HashGAN (class)

  • init (method)

    Defines the HashGAN network.

  • loss_D (method)

    Computes the dicriminator and encoder losses and stores them as attributes of HashGAN class.

  • loss_G (method)

    Computes the feature matching loss and stores it as an attribute of HashGAN class.

  • step_opt (method)

    Computes gradient and step optimizer.

  • generate_code_label (method)

    Generates codes using current encoder and labels of all datapoints in given dataloader.

  • eval (method)

    Evaluates the hashgan network with given query set and database set.

  • train (method)

    Train the network.

• Define the net, train, evaluate

utils.py

• get_trmat (function)

  Builds transform matrix to compute consistent bit loss.

• set_input (function)

  Makes input random variable consisting of a continuous part and a binary part of input random variable.

• get_prec_topn (function)

  Computes precision@topn with given query and database codes.

• bit_entropy (function)

  Computes the entropy of each bit of the given code.

models.py

• Generator (class)

  The definition of the model of generator network.

• Discriminator (class)

  The definition of the model of discriminator and encoder network.

References

• K.dizaji, F.Zheng, N.Nourabadi, Y.Yang, C.Deng, H.Huang. “Unsupervised Deep Generative Adversarial Hashing Network”, CVPR, 2018.