SNcGAN - Generate Conditional Images

Live demo: http://adeel.io/sncgan/

Spectral Norm + Conditional GAN

This is a hard fork of @minhnhat93's TensorFlow Spectral Normalization GAN implementation available at: https://github.com/minhnhat93/tf-SNDCGAN. A number of improvements have been made, including the addition of conditioning.

The GAN architecture is depicted here:

In order to train, you will need:

• Python 3.5+
• TensorFlow
• NumPy
• SciPy 0.19.1
• Pillow
• Pandas

To train, first download CelebA dataset and metadata from http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html:

• Unzip img_align_celeba.zip to /data/img_align_celeba
• Copy list_attr_celeba.txt to /data/img_align_celeba

To train, simply run:

• python train.py

As training progresses, check the output folder for generated samples over time. The first half of the samples grid should be males, and second half females.

To test after training is complete, run:

• python generate_images.py

You can edit generate_images.py, and modify the conditioning labels. For example, to get new samples of males, smiling, with black hair and mustaches, set the following flags in generate_images.py:

    sample_y['Male'] = 1
    sample_y['Smiling'] = 1
    sample_y['Black_Hair'] = 1
    sample_y['Blond_Hair'] = 0
    sample_y['Mustache'] = 1

Samples will be generated in the output folder, in collage.png.