Home

Awesome

BEGAN: Boundary Equilibrium Generative Adversarial Network

Basic Tensorflow implementation of BEGAN in <= 250 lines of code.

Dependencies

  1. Tensorflow r0.12 (will update to newest one next week)
  2. Tensorbayes
  3. Numpy
  4. Pillow

Train

Data Preparation

Download img_align_celeba.zip for the CelebA dataset and unzip to a directory. Then save the 64 x 64 CelebA crops in .mat format.

python prepare_celeba_zoom.py --source-dir /path/to/img_align_celeba --dest-file /path/to/celeba_64_zoom.mat

Train

To train the BEGAN model, execute

python main.py --data /path/to/celeba_64_zoom.mat

Results

Preliminary results after 140k iterations for config (gamma=0.5, embedding size=64) provided:

BEGAN image