VAE-GAN-CelebA

Python code related to the paper: "Reconstructing Faces from fMRI Patterns using Deep Generative Neural Networks" by VanRullen & Reddy (2019)

This folder contains:

• a link to a pre-trained VAE-GAN model checkpoint 'vaegan_celeba.ckpt' (~15 epochs on CelebA dataset=50,000 batches of 64 images)
• a set of .py functions for the VAE-GAN face decomposition/reconstruction model, in particular:
  • VAEGAN_image2latent.py => goes from any image file to the corresponding 1024D latent encoding (saved as a Matlab .mat file)
  • VAEGAN_latent2image.py => goes from a 1024D latent encoding (Matlab .mat file) to the corresponding image(s)
• (optional) a link to download the fMRI datasets (4 subjects, each saw > 8,000 faces in the scanner) and some Matlab analysis code

Example usage:

VAEGAN_image2latent.py -i example.jpg     #this will create example_z.mat with the 1024 latent vars
VAEGAN_latent2image.py -i example_z.mat   #this will generate example_z_g.jpg (and also example_z_g.mat)

Requirements:

• Python >= 3.4
• Tensorflow >= 1.8
• matplotlib, numpy, scipy, skimage