Tensorflow implementation of AmbientGAN

Tensorflow implementation of AmbientGAN on celebA dataset.
The implicit generative model learns to recover true underlying distribution given only lossy measurement models.

Network

Folder setting

-data
  -img_align_celeba
    -img1.jpg
    -img2.jpg
    -...

Measurement models

• block_pixel : Each pixel is independently set to zero with probability p.
• block_patch : A randomly chosen k × k patch is set to zero.
• keep_patch : All pixels outside a randomly chosen k × k patch are set to zero.
• conv_noise: k sized gaussian kernel is convolved and noise is added from the distribution Θ ∼ pθ.

Requirements

• python 2.7
• Tensorflow 1.4
• numpy
• cv2 (to save image tile)

Training

$ python train.py --measurement=block_pixel

To continue training

$ python train.py --measurement=block_pixel --continue_training=True

Block-Pixels

Trained CelebA images (pixels blocked with p=0.5)

Results

Block-Patch

Trained CelebA images (randomly blocked 32x32 patch)

Results

Keep-Patch

Trained CelebA images (randomly kept 32x32 patch)

Results

Convolve+Noise

Trained CelebA images (gaussian filter + noise(stddev=0.1))

ToDo

• Train Convolve+Noise