HelenMao / MSGAN

MSGAN: Mode Seeking Generative Adversarial Networks for Diverse Image Synthesis (CVPR2019)

Mode Seeking Generative Adversarial Networks for Diverse Image Synthesis

Pytorch implementation for our MSGAN (Miss-GAN). We propose a simple yet effective mode seeking regularization term that can be applied to arbitrary conditional generative adversarial networks in different tasks to alleviate the mode collapse issue and improve the diversity.

Contact: Qi Mao ([e-mail]), Hsin-Ying Lee ([e-mail]), and Hung-Yu Tseng ([e-mail])

Paper

Mode Seeking Generative Adversarial Networks for Diverse Image Synthesis
Qi Mao*, Hsin-Ying Lee*, Hung-Yu Tseng*, Siwei Ma, and Ming-Hsuan Yang
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019 (* equal contribution)
[arxiv]

Citing MSGAN

If you find MSGAN useful in your research, please consider citing:

@inproceedings{MSGAN,
  author = {Mao, Qi and Lee, Hsin-Ying and Tseng, Hung-Yu and Ma, Siwei and Yang, Ming-Hsuan},
  booktitle = {IEEE Conference on Computer Vision and Pattern Recognition},
  title = {Mode Seeking Generative Adversarial Networks for Diverse Image Synthesis},
  year = {2019}
}

Example Results

Usage

Prerequisites

• Python 3.5 or Python 3.6
• Pytorch 0.4.0 and torchvision (https://pytorch.org/)
• TensorboardX
• Tensorflow (for tensorboard usage)

Install

• Clone this repo:

git clone https://github.com/HelenMao/MSGAN.git

Training Examples

Download datasets for each task into the dataset folder

mkdir datasets

Conditoned on Label

• Dataset: CIFAR-10
• Baseline: DCGAN

cd MSGAN/DCGAN-Mode-Seeking
python train.py --dataroot ./datasets/Cifar10

Conditioned on Image

• Paired Data: facades and maps
• Baseline: Pix2Pix
  

You can download the facades and maps datasets from the BicycleGAN [Github Project].
We employ the network architecture of the BicycleGAN and follow the training process of Pix2Pix.

cd MSGAN/Pix2Pix-Mode-Seeking
python train.py --dataroot ./datasets/facades

• Unpaired Data: Yosemite (summer <-> winter) and Cat2Dog (cat <-> dog)
• Baseline: DRIT
  

You can download the datasets from the DRIT [Github Project].
Specify --concat 0 for Cat2Dog to handle large shape variation translation

cd MSGAN/DRIT-Mode-Seeking
python train.py --dataroot ./datasets/cat2dog

Conditioned on Text

• Dataset: CUB-200-2011
• Baseline: StackGAN++
  

You can download the datasets from the StackGAN++ [Github Project].

cd MSGAN/StackGAN++-Mode-Seeking
python main.py --cfg cfg/birds_3stages.yml

Pre-trained Models

• DCGAN-Mode-Seeking
• Pix2Pix-Mode-Seeking for facades and maps
• DRIT-Mode-Seeking for Yosemite and Cat2dog
• StackGAN++-Mode-Seeking

Download and save them into

./models/

Evaluation

For Pix2Pix, DRIT, and StackGAN++, please follow the instructions of corresponding github projects of the baseline frameworks for more evaluation details.

Testing Examples

DCGAN-Mode-Seeking

python test.py --dataroot ./datasets/Cifar10 --resume ./models/DCGAN-Mode-Seeking/00199.pth

Pix2Pix-Mode-Seeking

python test.py --dataroot ./datasets/facades --checkpoints_dir ./models/Pix2Pix-Mode-Seeking/facades --epoch 400

python test.py --dataroot ./datasets/maps --checkpoints_dir ./models/Pix2Pix-Mode-Seeking/maps --epoch 400

DRIT-Mode-Seeking

python test.py --dataroot ./datasets/yosemite --resume ./models/DRIT-Mode-Seeking/yosemite/01200.pth --concat 1

python test.py --dataroot ./datasets/cat2dog --resume ./models/DRIT-Mode-Seeking/cat2dog/01999.pth --concat 0

StackGAN++-Mode-Seeking

python main.py --cfg cfg/eval_birds.yml 

Reference

• DCGAN [Paper]
• Pix2Pix and BicycleGAN [Github Project]
• DRIT [Github Project]
• StackGAN++ [Github Project]

Quantitative Evaluation Metrics

• FID [Github Project]
• LPIPS [Github Project]
• NDB and JSD [Github Project]