Repository files navigation

This is a Tensorflow / Tensorlayer implementation of α-GAN for generating images to be used in EEG & fMRI deep image reconstruction.

α-GAN:Variational Approaches for Auto-Encoding Generative Adversarial Networks

Tensorflow - v1.8.0

Tensorlayer - v1.9.0

The training dataset must first be converted into a.tfrecord format.

This can be done by going toutils.py and modifyingclass_text_to_int(label) to contain the list of classes, and runningconvert_tfrecord(data_dir, save_dir, filename). An example is provided at the bottom ofutils.py which you can run by executingutils.py.

(data_dir should contain all the folders with the dataset labels, and all the dataset images should be in their respective folder)

Before training the α-GAN, make sure the directory paths inconfig.py correspond to the dataset locations.

Execute the training by running the following command

python3 main.py

This will train the α-GAN and save the model incheckpoints_dir every epoch.

Generator testing is split into two parts: training set, and generation performance. These two are saved insave_gan_dir andsave_test_gan_dir respectively.

This extracts the features from the given folder of images using the trained encoder, and stores them inencoded_feat.pkl.

python3 main.py --mode=encode

This reconstructs the folder of images from the encoding section by using the extracted features fromencoded_feat.pkl to generate images.

python3 main.py --mode=genpython3 main.py --mode=generate