Image Generation Deep Learning

This repository contains a deep learning system that performs image generation using a deep convolutional generative adversarial network (DCGAN). The dataset url and the associated article url are provided below.

Dataset url: https://www.dropbox.com/sh/isslk5zkp9ekqtc/AABTfbuLYRID6NhDvq1Vi7Hha?dl=0

Article url: https://medium.com/@yvanscher/using-gans-to-create-monsters-for-your-game-c1a3ece2f0a0

Installation

It is recommended to install conda and then create an environment for the system using the environment.yaml file. A suggestion on how to install the system and activate the environment is provided below.

git clone https://github.com/anpenta/image-generation-deep-learning.git
cd image-generation-deep-learning
conda env create -f environment.yaml
conda activate image-generation-deep-learning

Running the system

To run the system for training you can provide commands through the terminal using the train module. An example is given below.

python3 train.py ./pokemon-data 1000 8 100 ./output

This will train a model using data from the ./pokemon-data directory for 1000 epochs with a batch size of 8 and save the training plots and the trained model every 100 epochs in the ./output directory. An example of how to see the parameters for training is provided below.

python3 train.py --help

To run the system for generating images you can provide commands through the terminal using the generate module. An example is given below.

python3 generate.py ./generator.h5 50 ./output

This will use the generator from the ./generator.h5 file to generate 50 images and save them in the ./output directory. An example of how to see the parameters for generation is provided below.

python3 generate.py --help

Results

As an example, below are the training results we get after training a model for 1460 epochs. The generator does not generate sensible images every time, so I chose five of the best ones. In general, the training is not smooth and you may have to restart training if the discriminator outperforms the generator for many consecutive epochs or vice versa.

Sources

• Radford, Alec, Luke Metz, and Soumith Chintala. "Unsupervised representation learning with deep convolutional generative adversarial networks." arXiv preprint arXiv:1511.06434 (2015).