One GPU, Learning In Less Than A Day! Ultra-Fast Learning GAN, "Lightweight GAN"

3 main points
✔️ Skip-Layer Excitation and self-supervised Discriminator proposed to reduce parameters significantly 
✔️ It is possible to train even with small amounts of data
✔️ Even a 1024×1024 image can be trained on a single GPU in a few hours

Towards Faster and Stabilized GAN Training for High-fidelity Few-shot Image Synthesis
written by Anonymous
(Submitted on 29 Sep 2020)
Comments: Accepted at ICLR2021
Subjects: Computer Vision and Pattern Recognition (cs.CV); Image and Video Processing (eess.IV)
   Comm 

overview

Until now, GANs, especially those like StyleGANs, have required a considerable amount of time for training. Our new lightweight GAN introduces two modules, Skip-Layer Excitation and self-supervised Discriminator, which significantly reduces the number of parameters. As a result, we are able to train at an astonishingly low cost: high quality images, small amounts of data, one GPU, and a few hours!

Here are the resulting images shown at the beginning. Both images are 1024×1024 in size and are trained with one GPU and 1000 images. The training time is 20 hours for the landscape image and 10 hours for the face image.

けやみぃ

I am a first-year student at the Faculty of Engineering, Kyoto University, and I am interested in image generation and image transformation using GAN.