liannice/aguit

Pytorch Implement of Paper ‘’Attribute Guided Unpaired Image-to-image Translation with Semi-supervised Learning"

AGUIT: Attribute Guided Unpaired Image Translation

Code will be available soon.

Project | Paper | Demo (Coming Soon)

"Attribute Guided Unpaired Image-to-Image Translation with Semi-supervised Learning." Li, X., Hu, J., Zhang, S., Hong, X., Ye, Q., Wu, C., & Ji, R. arXiv preprint.

Merits

• Unpaired: Or Unsupervised in many works. The models are trained by using unpaired data in this setting. However, we prefer to term this as Unpaired instead of Unsupervised as the label of image domain (e.g., hair color, smile or not) is still needed.

• Semi-Supervised: An attempt to reduce the demand of such domain-labeled (or attribute-labeled) data in a semi-supervised setting. It means that some images in the training set are labeled and some are not.

• Multi-Modal & Multi-Domain: One/Many-to-Many translation.

• Single Model: All One/Many-to-Many translations are achieved by one single model.

• Disentanglement & Fine Control: The encoded attribute/domain representations are disentangled and can be fine controled.

• Disentangled Transfer: We introduce this new challenge for image-to-image translation community, which adopts the disentangled representation to translate data less related with the training set. For example, the models trained with human faces can be used to put eyeglasses on the dog's or cat's faces in the disentangled transfer.

Usage

You need to use python3 to run this code. Other requirements are:

• pytorch

• tensorboardX

Besides, you need to format your dataset, like the ./datasets/CelebA_example. (the labeled folder and label.txt consist of the labeled images and their labels, and the unlabeled folder consists of the unlabeled image.)

Run the code with default configuration by (python train.py --config configs/face.yaml).

The outputs and checkpoint would be in ./outputs

And test your model by modify the code in test.py and (python test.py --config configs/face.yaml --checkpoint #your_checkpoint --input #your_input --output_dir #your_output_dir)

The gui-test-code is waiting for a final check, it would be more easy for you to test your model in the future. Please wait for couples of days.

Citation

If our paper helps your research, please cite it in your publications:

@article{li2019attribute,
  title={Attribute Guided Unpaired Image-to-Image Translation with Semi-supervised Learning},
  author={Li, Xinyang and Hu, Jie and Zhang, Shengchuan and Hong, Xiaopeng and Ye, Qixiang and Wu, Chenglin and Ji, Rongrong},
  journal={arXiv preprint arXiv:1904.12428},
  year={2019}
}

If you have any problem, please feel free to contact us.

Related Work

Our work benefits from code of MUNIT, DRIT, StarGAN, etc.