Slide 1

Slide 1 text

1 25-27 NOVEMBER SOFTWARE TESTING, MACHINE LEARNING AND COMPLEX PROCESS ANALYSIS Unpaired Image-to-Image Translation using Transformer-based CycleGAN Chongyu Gu, Maxim Gromov

Slide 2

Slide 2 text

2 Motivation Convolutional Layer Vision Transformer [Dosovitskiy, Alexey, et al., ICLR, 2020] Ref: https://www.ibm.com/cloud/learn/convolutional- neural-networks

Slide 3

Slide 3 text

3 Motivation Visual results produced by TransGAN Unconditional image generation results by TransGAN [Yifan Jiang, Shiyu Chang, Zhangyang Wang, NeurIPS, 2021]

Slide 4

Slide 4 text

4 4 The Generator and Discriminator Networks The pipeline of the pure transform-based generator and discriminator of TransCycleGAN.

Slide 5

Slide 5 text

5 5 Architecture configuration of generator. Architecture configuration of discriminator. The Generator and Discriminator Networks

Slide 6

Slide 6 text

6 6 Main Results Samples of the horse2zebra 64 ×64 Samples of the zebra2horse 64 ×64 Our model reaches FID of 80.54 on horse2zebra 64 ×64 and 93.05 FID on zebra2horse 64 ×64.

Slide 7

Slide 7 text

7 7 Conclusion and Limitation We have introduced TransCycleGAN, the first pure transformer-based GAN for the task of image-to-image translation. Our experiments on the horse2zebra 64 × 64 benchmark demonstrate that the great potential of our new architecture. TransCycleGAN still has much room for exploration, such as going towards high-resolution translation tasks (e.g.,256 × 256) and experimenting on more datasets like Apple↔Orange, Summer↔Winter Yosemite, and Photo↔Art for style transfer, which is our future directions.

Slide 8

Slide 8 text

8 Thank you very much for your attention!