Ha0Tang/DAGAN

Contents

• Semantic Image Synthesis with DAGAN
• Installation
• Dataset Preparation
• Generating Images Using Pretrained Model
• Train and Test New Models
• Evaluation
• Acknowledgments
• Related Projects
• Citation
• Contributions
• Collaborations

Semantic Image Synthesis with DAGAN

Dual Attention GANs for Semantic Image Synthesis
Hao Tang¹, Song Bai², Nicu Sebe¹³.
¹University of Trento, Italy, ²University of Oxford, UK, ³Huawei Research Ireland, Ireland.
In ACM MM 2020.
The repository offers the official implementation of our paper in PyTorch.

In the meantime, check out our related CVPR 2020 paper Local Class-Specific and Global Image-Level Generative Adversarial Networks for Semantic-Guided Scene Generation, and TIP 2021 paper Layout-to-Image Translation with Double Pooling Generative Adversarial Networks.

Framework

Results of Generated Images

Cityscapes (512×256)

Facades (1024×1024)

ADE20K (256×256)

CelebAMask-HQ (512×512)

Results of Generated Segmenation Maps

License

Copyright (C) 2020 University of Trento, Italy.

All rights reserved. Licensed under the CC BY-NC-SA 4.0 (Attribution-NonCommercial-ShareAlike 4.0 International)

The code is released for academic research use only. For commercial use, please contact [email protected].

Installation

Clone this repo.

git clone https://github.com/Ha0Tang/DAGAN
cd DAGAN/

This code requires PyTorch 1.0 and python 3+. Please install dependencies by

pip install -r requirements.txt

This code also requires the Synchronized-BatchNorm-PyTorch rep.

cd DAGAN_v1/
cd models/networks/
git clone https://github.com/vacancy/Synchronized-BatchNorm-PyTorch
cp -rf Synchronized-BatchNorm-PyTorch/sync_batchnorm .
cd ../../

To reproduce the results reported in the paper, you would need an NVIDIA DGX1 machine with 8 V100 GPUs.

Dataset Preparation

Please download the datasets on the respective webpages.

• Facades: 55.8M, here.
• DeepFashion: 592.3M, here.
• CelebAMask-HQ: 2.7G, here.
• Cityscapes: 8.4G, here.
• ADE20K: 953.7M, here.
• COCO-Stuff: 21.5G, here.

We also provide the prepared datasets for your convience.

sh datasets/download_dagan_dataset.sh [dataset]

where [dataset] can be one of facades, deepfashion, celeba, cityscapes, ade20k, or coco_stuff.

Generating Images Using Pretrained Model

1. Download the pretrained models using the following script,

sh scripts/download_dagan_model.sh GauGAN_DAGAN_[dataset]

where [dataset] can be one of cityscapes, ade, facades, or celeba.

2. Change several parameter and then generate images using test_[dataset].sh. If you are running on CPU mode, append --gpu_ids -1.
3. The outputs images are stored at ./results/[type]_pretrained/ by default. You can view them using the autogenerated HTML file in the directory.

Train and Test New Models

1. Prepare dataset.
2. Change several parameters and then run train_[dataset].sh for training. There are many options you can specify. To specify the number of GPUs to utilize, use --gpu_ids. If you want to use the second and third GPUs for example, use --gpu_ids 1,2.
3. Testing is similar to testing pretrained models. Use --results_dir to specify the output directory. --how_many will specify the maximum number of images to generate. By default, it loads the latest checkpoint. It can be changed using --which_epoch.

Evaluation

• FID: mseitzer/pytorch-fid
• FRD: Ha0Tang/GestureGAN
• LPIPS: richzhang/PerceptualSimilarity
• DRN: fyu/drn [model: drn-d-105_ms_cityscapes.pth]
• UperNet: CSAILVision/semantic-segmentation-pytorch [model: baseline-resnet101-upernet]
• DeepLab: kazuto1011/deeplab-pytorch [model: deeplabv2_resnet101_msc-cocostuff164k-100000.pth]

For more details, please refer to this issue.

Acknowledgments

This source code is inspired by both GauGAN/SPADE and LGGAN.

Related Projects

ECGAN | LGGAN | SelectionGAN | DPGAN | PanoGAN | Guided-I2I-Translation-Papers

Citation

If you use this code for your research, please consider giving stars ⭐ and citing our papers 🦖:

DAGAN

@inproceedings{tang2020dual,
  title={Dual Attention GANs for Semantic Image Synthesis},
  author={Tang, Hao and Bai, Song and Sebe, Nicu},
  booktitle ={ACM MM},
  year={2020}
}

ECGAN

@article{tang2023edge,
  title={Edge Guided GANs with Contrastive Learning for Semantic Image Synthesis},
  author={Tang, Hao and Qi, Xiaojuan and Sun, Guolei, and Xu, Dan and and Sebe, Nicu and Timofte, Radu and Van Gool, Luc},
  journal={ICLR},
  year={2023}
}

LGGAN

@article{tang2022local,
  title={Local and Global GANs with Semantic-Aware Upsampling for Image Generation},
  author={Tang, Hao and Shao, Ling and Torr, Philip HS and Sebe, Nicu},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)},
  year={2022}
}

@inproceedings{tang2019local,
  title={Local Class-Specific and Global Image-Level Generative Adversarial Networks for Semantic-Guided Scene Generation},
  author={Tang, Hao and Xu, Dan and Yan, Yan and Torr, Philip HS and Sebe, Nicu},
  booktitle={CVPR},
  year={2020}
}

SelectionGAN

@article{tang2022multi,
  title={Multi-Channel Attention Selection GANs for Guided Image-to-Image Translation},
  author={Tang, Hao and Torr, Philip HS and Sebe, Nicu},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)},
  year={2022}
}

@inproceedings{tang2019multi,
  title={Multi-channel attention selection gan with cascaded semantic guidance for cross-view image translation},
  author={Tang, Hao and Xu, Dan and Sebe, Nicu and Wang, Yanzhi and Corso, Jason J and Yan, Yan},
  booktitle={CVPR},
  year={2019}
}

DPGAN

@article{tang2021layout,
  title={Layout-to-image translation with double pooling generative adversarial networks},
  author={Tang, Hao and Sebe, Nicu},
  journal={IEEE Transactions on Image Processing (TIP)},
  volume={30},
  pages={7903--7913},
  year={2021}
}

PanoGAN

@article{wu2022cross,
  title={Cross-View Panorama Image Synthesis},
  author={Wu, Songsong and Tang, Hao and Jing, Xiao-Yuan and Zhao, Haifeng and Qian, Jianjun and Sebe, Nicu and Yan, Yan},
  journal={IEEE Transactions on Multimedia (TMM)},
  year={2022}
}

Contributions

If you have any questions/comments/bug reports, feel free to open a github issue or pull a request or e-mail to the author Hao Tang ([email protected]).

Collaborations

I'm always interested in meeting new people and hearing about potential collaborations. If you'd like to work together or get in contact with me, please email [email protected]. Some of our projects are listed here.

Take a few minutes to appreciate what you have and how far you've come.