FIFA VAE

This repository contains an implementation of a Variational Autoencoder for creating fake FIFA football players. See in depth tutorial on Medium post titled Generating fake FIFA football players with Variational Autoencoders and Tensorflow.

File list

• donwloader.py: FIFA images script downloader
• main.py: Script to train the network (see parameters below)
• network.py: Class with VAE network definition
• dataset.py: Data pipeline using tensorflow dataset API
• utils.py: Simple functions that are used in multiple files
• eval.py: Script to evaluate network after training (see parameters below)

Downloader

The downloader script is downloader.py and it can receive two parameters:

--n_threads                  Number of threads to create 
--destination_folder         Directory were data will be downloaded to

Run downloader

$ python downloader.py  20  Data/Images

Training

Once the data has been downloaded, the training phase can start. The training file is main.py and receives the following parameters:

-- learning_rate      Initial learning rate
--epochs              Number of training epochs 
--batch_size          Minibatch training size
--latent_dim          Number of latent space dimensions
--test_image_number   Number of test images to recover during training
--epochs_to_plot      Number of epochs before saving test sample of reconstructed images
--save_after_n        Number of epochs before saving network
--logdir              Log folder
--inputs_decoder      Size of decoder input layer
--data_path           Data folder path
--shuffle             Shuffle dataset (boolean)

Run Training

$ python main.py --latent_dim=2 --batch_size=512 --data_path=Data/Images --epochs=1000

Examples

The examples can only being executed once the training it's complete or when we have some previously saved model.

--learning_rate             Learning rate used in trained (model names use this parameter)
--max_epoch                 Max epoch saved 
--latent_dim                Number of latent space dimensions for model
--create_fake_players       Number of fake players to generate
--recover_top_5             Reconstructs top 5 FIFA players from their latent vector
--meshgrid                  Creates meshgrid of points in latent space and shows reconstructed images [only when latent_dim = 2]
--mix_players               List with two image paths to mix
--n_interpolations          Number of interpolations between mixed players
--countries                 List of countries to compute average player

Run examples

$ python eval.py --latent_dim=15 --max_epoch=980

Some examples

Compare original images with the reconstructed ones.

When our model was trained with a two dimensional latent space, we can get the following plot.

Compute the interpolation between two different players.

Generate average players for each of the given countries.