StarGAN-Tensorflow

Implementation of StarGAN in Tensorflow

StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation Official Pytorch Code

This code is mainly revised from goldkim92's code base on official pytorch code.

• Modifying the code to be more consistent to the official implementation.
• Fixing the bug in lost calculation.
• More testing function added.
• Adding Residual Block (base on this code)

Prerequisites

• Python 3.5
• Tensorflow 1.3.0
• Scipy
• tqdm

Usage

Only CelebA part is implemented.

First, download dataset with:

$ python download.py

To train a model:

$ python main.py --phase=train --image_size=64 --batch_size=16

The default classification method is using sigmoid. If the attributes you chose satisfy sigle attribute classification (ex: hair color only. Or if you can access to RAFD), you could also try softmax.

$ python main.py --phase=train --image_size=64 --batch_size=16 --c_method=Softmax

The default adversarial training method is improved WGAN. You could also try different method such as LSGAN or GAN. But personally I've only tried the improved WGAN.

$ python main.py --phase=train --image_size=64 --batch_size=16 --adv_type=LSGAN

The output format of the sample image during training:

To test a model by given a specific attribute:

$ python main.py --phase=test --image_size=64 --binary_attrs=100000

The output format of the image is like:

Bianry attributes are now set up with the following sequence:

'Black_Hair', 'Blond_Hair', 'Brown_Hair', 'Male', 'Young', 'Pale_Skin'

You could modify the attributes in the main.py

Sample 100 images from testing data and test each image with each attribute:

$ python main.py --phase=test_all --image_size=64

The output format of the image is like:

To test the classifier of a model:

$ python main.py --phase=aux_test --image_size=64

Result