Pytorch implementation for learning to synthesize images in a hierarchical, stagewise manner by disentangling background, object shape and object appearance.

Krishna Kumar Singh*, Utkarsh Ojha*, Yong Jae Lee
project | arxiv | demo video | talk video
CVPR 2019 (Oral Presentation)

• Linux
• Python 2.7
• Pytorch 0.4.1
• TensorboardX 1.2
• NVIDIA GPU + CUDA CuDNN

git clone https://github.com/kkanshul/finegan
cd finegan

Note: You only need to download the data if you wish to train your own model.

Download the formatted CUB data from this link and extract it inside the data directory

cd data
unzip birds.zip
cd ..

Pretrained generator models for CUB, Stanford Dogs are available at this link. Download and extract them in the models directory.

cd models
unzip netG.zip
cd ../code/

In cfg/eval.yml:

• Specify the model path in TRAIN.NET_G.
• Specify the output directory to save the generated images in SAVE_DIR.
• Specify the number of super and fine-grained categories in SUPER_CATEGORIES and FINE_GRAINED_CATEGORIES according to our paper.
• Specify the option for using 'tied' latent codes in TIED_CODES:
  • if True, specify the child code in TEST_CHILD_CLASS. The background and parent codes are derived through the child code in this case.
  • if False, i.e. no relationship between parent, child or background code, specify each of them in TEST_PARENT_CLASS, TEST_CHILD_CLASS and TEST_BACKGROUND_CLASS respectively.
• Run python main.py --cfg cfg/eval.yml --gpu 0

In cfg/train.yml:

• Specify the dataset location in DATA_DIR.
  • NOTE: If you wish to train this on your own (different) dataset, please make sure it is formatted in a way similar to the CUB dataset that we've provided.
• Specify the number of super and fine-grained categories that you wish for FineGAN to discover, in SUPER_CATEGORIES and FINE_GRAINED_CATEGORIES.
• Specify the training hyperparameters in TRAIN.
• Run python main.py --cfg cfg/train.yml --gpu 0

If you find this code useful in your research, consider citing our work:

@inproceedings{singh-cvpr2019,
  title = {FineGAN: Unsupervised Hierarchical Disentanglement for Fine-Grained Object Generation and Discovery},
  author = {Krishna Kumar Singh and Utkarsh Ojha and Yong Jae Lee},
  booktitle = {CVPR},
  year = {2019}
}

We thank the authors of StackGAN++: Realistic Image Synthesis with Stacked Generative Adversarial Networks for releasing their source code.

For any questions regarding our paper or code, contact Krishna Kumar Singh and Utkarsh Ojha.