GANs N' ROSES

Uses a Deep Convolutional Generative Adversial Network to generate images of roses using tensorflow.

The main obejctive of this repo is to provide users with a foundation to experiment on GANs without worrying about saving models, generated images and using Tensorboard. All the functions have been documented, instructions to get started on the theory and the explaination is provided here.

Installing the dependencies

Install virtualenv and creating a new virtual environment:

pip install virtualenv
virtualenv -p /usr/bin/python3 gnr

Install dependencies

pip3 install -r requirements.txt

Note:

• Install tensorflow using the following link.
• It is recomended that you use the GPU during training.
• The code works on tensorflow r0.12, change the parameter name from targets to labels at lines 123, 124 and 125 in the main.py file for version > r0.12

Dataset

The roses dataset was downloaded from google images using the Chrome ImageSpark plugin.

The dataset that I have used can be downloaded from this link.

• Extract the images into Dataset/Roses.

Note:

• Images other than roses and those with different shapes can be copied into the Roses directory.

Training the model

mission_control.py

This file contains the dataset location and the hyperparameters which are used by the model.

Note:

• IMAGE_SIZE : currently buggy don't change it from 64. The Generator model has to be modified in order to change the image size at the output of the generator. IMAGE_SIZE only changes the size of the real images feed to the Discriminator.

Training

Set the hyperparameters to the required values in the mission_control.py file and run the following:

python3 main.py

• Each run generates a new directory under Results/roses with the following format:

  <time_stamp>_<Z_DIM>_<BATCH_SIZE>_<N_ITERATIONS>_<LEARNING_RATE>_<BETA_1>

• This trains the model displaying the generated images after every 200 iternations in the Results/roses/<time_stamp...>/Generated_Images directory.

• The Discriminator and the Generator loss along with the Discriminator accuracy variations are plotted on Tensorboard which can be seen by running the code below.

  tensorboard --logdir="Results/roses/<time_stamp...>/Tensorboard"

Examples of generated images after 30000 iterations:

The example images were generated using the default hyperparameters present in the mission_control.py file.

Example 1

Example 2

GANs N' Roses Training

Thank You

If you find this repository helpful please share it and feel free to suggest any changes.