StarGAN

Description

The Keras impletation of the StarGAN.

Example

Image Dataset

CelebA Image Dataset

Version

Tensorflow:1.14.0

Keras:2.2.5

Note for the code

• list_attr_celeba.txt: The attribute file, containing one-hot encoding for each image. The dimension of label is 40.
• Image data used : in the dataset folder: CelebA\Img\img_align_celeba
• The training is 40 dims label. If only a few labels are considered, for example, 5 labels: ['Black_Hair','Blond_Hair','Brown_Hair','Male','Young'],then one solusion is to train the 40-label network at first, then concatenate extensions(5 input - 40 output, and 40 input - 5 output) on the pretrained model and train both pretrained model and extensions. The training with pretrained model is fast, taking less than 1 epoch.

Reference paper

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