Prinsphield / GeneGAN

GeneGAN: Learning Object Transfiguration and Attribute Subspace from Unpaired Data
https://arxiv.org/abs/1705.04932v1
GNU General Public License v3.0
144 stars 34 forks source link
attributes-subspace celeba-dataset generative-adversarial-network image-interpolation image-manipulation

GeneGAN: Learning Object Transfiguration and Attribute Subspace from Unpaired Data

By Shuchang Zhou, Taihong Xiao, Yi Yang, Dieqiao Feng, Qinyao He, Weiran He

If you use this code for your research, please cite our paper:

@inproceedings{DBLP:conf/bmvc/ZhouXYFHH17,
  author    = {Shuchang Zhou and
               Taihong Xiao and
               Yi Yang and
               Dieqiao Feng and
               Qinyao He and
               Weiran He},
  title     = {GeneGAN: Learning Object Transfiguration and Attribute Subspace from Unpaired Data},
  booktitle = {Proceedings of the British Machine Vision Conference (BMVC)},
  year      = {2017},
  url       = {http://arxiv.org/abs/1705.04932},
  timestamp = {http://dblp.uni-trier.de/rec/bib/journals/corr/ZhouXYFHH17},
  bibsource = {dblp computer science bibliography, http://dblp.org}
}

We have two following papers, DNA-GAN and ELEGANT, that generalize the method into multiple attributes case. It is worth mentioning that ELEGANT can transfer multiple face attributes on high resolution images. Please pay attention to our new methods!

Introduction

This is the official source code for the paper GeneGAN: Learning Object Transfiguration and Attribute Subspace from Unpaired Data. All the experiments are initially done in our proprietary deep learning framework. For convenience, we reproduce the results using TensorFlow.

cross


GeneGAN is a deterministic conditional generative model that can learn to disentangle the object features from other factors in feature space from weak supervised 0/1 labeling of training data. It allows fine-grained control of generated images on one certain attribute in a continous way.

Requirement

Training GeneGAN on celebA dataset

  1. Download celebA dataset and unzip it into datasets directory. There are various source providers for CelebA datasets. To ensure that the size of downloaded images is correct, please run identify datasets/celebA/data/000001.jpg. The size should be 409 x 687 if you are using the same dataset. Besides, please ensure that you have the following directory tree structure.
├── datasets
│   └── celebA
│       ├── data
│       ├── list_attr_celeba.txt
│       └── list_landmarks_celeba.txt
  1. Run python preprocess.py. It will take several miniutes to preprocess all face images. A new directory datasets/celebA/align_5p will be created.

  2. Run python train.py -a Bangs -g 0 to train GeneGAN on the attribute Bangs. You can train GeneGAN on other attributes as well. All available attribute names are listed in the list_attr_celeba.txt file.

  3. Run tensorboard --logdir='./' --port 6006 to watch your training process.

Testing

We provide three kinds of mode for test. Run python test.py -h for detailed help. The following example is running on our GeneGAN model trained on the attribute Bangs. Have fun!

1. Swapping of Attributes

You can easily add the bangs of one person to another person without bangs by running

python test.py -m swap -i datasets/celebA/align_5p/182929.jpg -t datasets/celebA/align_5p/022344.jpg
input target out1 out2
Swap Attribute


2. Linear Interpolation of Image Attributes

Besides, we can control to which extent the bangs style is added to your input image through linear interpolation of image attribute. Run the following code.

python test.py -m interpolation -i datasets/celebA/align_5p/182929.jpg -t datasets/celebA/align_5p/035460.jpg -n 5
interpolation target
Linear Interpolation


3. Matrix Interpolation in Attribute Subspace

We can do something cooler. Given four images with bangs attributes at hand, we can observe the gradual change process of our input images with a mixing of difference bangs style.

python test.py -m matrix -i datasets/celebA/align_5p/182929.jpg --targets datasets/celebA/align_5p/035460.jpg datasets/celebA/align_5p/035451.jpg datasets/celebA/align_5p/035463.jpg datasets/celebA/align_5p/035474.jpg -s 5 5
matrix
Matrix Interpolation