GAN - Pokémons

This project focuses on generating new Pokémon images by using convolutional neural networks as a generator and discriminator. The GAN is trained with the original Pokémon dataset.
For simplicity, the alpha channel was omitted so that only RGB images with a size of 128 x 128 were considered.

Remark: The png images must be saved in the data folder as pokemon_dataset.

The dataset has the following properties:

• Total amount: 819 images
• Size: 41 MB
• MIME-Type: png
• Shape: 256 x 256 x 4

Resources

FID-Score
FID-Implementation
Generative Adversarial Nets
Differentiable Augmentation for Data-Efficient GAN Training

Usage

Installing Packages

For the packages use miniconda. See Miniconda.

Use the Anaconda Prompt terminal for the following steps:

conda create --name myenv python=3.8.5 pip

When conda asks you to proceed proceed ([y]/n)?, type y:

Activate the new environment:

conda activate myenv

Install pytorch

conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch

Install the packages via requirements.txt

 pip install -r requirements.txt

Preprocessing images

The images will be resized to 128 x 128 x 3 at the beginning

python -m pokemon.resize_images

Train new generator

A trained model is available underpokemon\trained models. If you want to train a new generator, you have to execute this command

python -m pokemon.generator.train

The hyperparameters are in the "train_config.json" file.
The trained models can be found in the folder pokemon\trained_models

Generate images

For generating new Pokémon images, you can choose between two modules.
For single images

python -m pokemon.generate_images

For grid images

python -m pokemon.generate_grid_images

The images can be found under the following paths:

• pokemon\data\generated_images,
• pokemon\data\generated_grid_images.

Evaluation

The FID-Score will be used to evalute the GAN.
It is important to create the same number of images as the original data.
The following folders are required:

• Images: pokemon\data\pokemon_preprocessed, with the size of 819 images,
• New Images: pokemon\data\generated_images, with the size of 819 images.

To evaluate the Pokémon images, you have to run

python -m pokemon.evaluation.pytorch-fid.src.pytorch_fid.fid_score --num-workers 2 pokemon\data\generated_images pokemon\data\pokemon_preprocessed

Model	FID-Score
2022-03-08_15_49_59.pt	173.13