Hierarchically-nested Adversarial Network (Pytorch implementation)

We call our method HDGAN, referring to High-Definition results and the idea of Hierarchically-nested Discriminators

Zizhao Zhang*, Yuanpu Xie*, Lin Yang, "Photographic Text-to-Image Synthesis with a Hierarchically-nested Adversarial Network", CVPR (2018) * indicates contribution

Visual results (Left: compared against StackGAN; Right: multi-resolution generator outputs)

Dependencies

• Python 3
• Pytorch 0.3.1
• Anaconda 3.6
• Tensorflow 1.4.1 (for evaluation only)

Data

Download preprocessed data in /Data.

• Download birds to Data/birds
• Download flowers Data/flowers
• Download coco Data/coco. Also download COCO images to Data/coco/coco_official

Training

• For bird: goto train/train_gan: device=0 sh train_birds.sh
• For flower: goto train/train_gan: device=0 sh train_flower.sh
• For coco: goto train/train_gan: device=0,1 sh train_coco.sh

To use multiple GPUs, simply set device='0,1,..' as a set of gpu ids.

Monitor your training in two ways

• Launch Visdom (see here): python -m visdom.server -port 43426 (keep the same port id with _port defined in plot_utils.py). Then access http://localhost:43426 from the browser.
• Check fixed sample results per epoch in the checkpoint folder.

Testing

• Go to test/test_gan:
• sh test_birds.sh for birds
• sh test_flowers.sh for flower
• sh test_coco.sh for coco

Evaluation

We provide multiple evaluation tools to ease test. Evaluation needs the sampled results obtained in Testing and saved in ./Results.

• Go to /Evaluation

Inception score

• Download inception models to the Evaluation/inception_score/inception_finetuned_models folder compute inception scores.
• Compute the inception score: sh compute_inception_score.sh

MS-SSIM

• Compute the MS-SSIM score: sh compute_ms_ssim.sh

VS-Similarity

• Download models to the Evaluation/neudist/neudist_[dataset] folder.
• Evaluate the VS-smilarity score: sh compute_neudist_score.sh

Pretrained Models

We provide pretrained models for birds, flowers, and coco.

• Download the pretrained models. Save them to the Models/ folder.
• It contains HDGAN for birds and flowers, visual similarity model for birds and flowers

Acknowlegements

• StakGAN Tensorflow implementation
• MS-SSIM Python implementation
• Inception score Tensorfow implementation
• Process COCO data Torch implementation

Citation

If you find HDGAN useful in your research, please cite:

@inproceedings{zhang2018hdgan,
Author = {Zizhao Zhang and Yuanpu Xie and Lin Yang},
Title = {Photographic Text-to-Image Synthesis with a Hierarchically-nested Adversarial Network},
Year = {2018},
booktitle = {CVPR},
}

License

MIT