XiaozhuFang / InvolutionGAN

The new involution operator was implemented in GAN, compared with DCGAN and SAGAN.

Involution GAN

Run the main.py and set parameters.py

• you can download
  • MNIST dataset: http://yann.lecun.com/exdb/mnist/
  • CelebA dataset: http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html

Acknowledgement

@InProceedings{Li_2021_CVPR,
    author = {Li, Duo and Hu, Jie and Wang, Changhu and Li, Xiangtai and She, Qi and Zhu, Lei and Zhang, Tong and Chen, Qifeng},
    title = {Involution: Inverting the Inherence of Convolution for Visual Recognition},
    booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month = {June},
    year = {2021}
}

We thanks to the code from

• https://github.com/heykeetae/Self-Attention-GAN
• https://github.com/ChristophReich1996/Involution
• https://github.com/mseitzer/pytorch-fid

Network Structure

Sample

imgaes of iGAN have more saturated and smoother color over SAGAN and DCGAN

Experiment Result

To distinguish the performance of iGAN, DCGAN, and SAGAN, we ensure the other parts of networks consistent as much as possible. The simplest DCGAN is the basic structure, and SAGAN adds an additional self-attention layer in both generator and discriminator. Likewise, iGAN adds an additional involution at the same location.

Hyperparameter tuning: different cases with respect to involution kernel size, group of involution, and reduction ratio.

Report

paper_33.pdf

Slides

slides_33.pdf