GAN Compression: Efficient Architectures for Interactive Conditional - - PowerPoint PPT Presentation

▶
gan compression efficient architectures for interactive
SMART_READER_LITE
LIVE PREVIEW

GAN Compression: Efficient Architectures for Interactive Conditional - - PowerPoint PPT Presentation

Jan 25, 2024 277 likes •444 views

GAN Compression: Efficient Architectures for Interactive Conditional GANs Muyang Li 1,3 , Ji Lin 1 , Yaoyao Ding 1,3 , Zhijian Liu 1 , Jun-Yan Zhu 2 , and Song Han 1 1 Massachusetts Institute of Technology 2 Adobe Research. 3 Shanghai Jiao Tong

slide-1
SLIDE 1

1Massachusetts Institute of Technology 2Adobe Research. 3Shanghai Jiao Tong University

Muyang Li1,3, Ji Lin1, Yaoyao Ding1,3, Zhijian Liu1, Jun-Yan Zhu2, and Song Han1

GAN Compression: Efficient Architectures for Interactive Conditional GANs

1

slide-2
SLIDE 2

2

Conditional GANs are Computationally-Intensive.

slide-3
SLIDE 3

2

Pix2pix, CVPR’17

Conditional GANs are Computationally-Intensive.

slide-4
SLIDE 4

2

MACs

MobileNet ResNet-50 CycleGAN GauGAN

Pix2pix, CVPR’17

Conditional GANs are Computationally-Intensive.

slide-5
SLIDE 5

2

MACs

281G 57G 4.0G 0.5G MobileNet ResNet-50 CycleGAN GauGAN

GANs requires 10×-500× more computations than image recognition models!

Pix2pix, CVPR’17

Conditional GANs are Computationally-Intensive.

500x

slide-6
SLIDE 6

3

General-purpose Compression Framework—GAN Compression

Distillation + Channel Pruning + NAS

slide-7
SLIDE 7

4

GAN Compression Reduces the Computation by 9-21x

Horse→zebra CycleGAN, Zhu et al. Edge→shoes Pix2pix, Isola et al. Cityscapes GauGAN, Park et al.

slide-8
SLIDE 8

4

GAN Compression Reduces the Computation by 9-21x

MACs

1G 10G 100G 1000G

Original Model GAN Compression

Horse→zebra CycleGAN, Zhu et al. Edge→shoes Pix2pix, Isola et al. Cityscapes GauGAN, Park et al.

slide-9
SLIDE 9

4

GAN Compression Reduces the Computation by 9-21x

MACs

1G 10G 100G 1000G

Original Model GAN Compression

2.7G 57G

Horse→zebra CycleGAN, Zhu et al. Edge→shoes Pix2pix, Isola et al. Cityscapes GauGAN, Park et al.

slide-10
SLIDE 10

4

GAN Compression Reduces the Computation by 9-21x

MACs

1G 10G 100G 1000G

Original Model GAN Compression

4.8G 2.7G 57G 57G

Horse→zebra CycleGAN, Zhu et al. Edge→shoes Pix2pix, Isola et al. Cityscapes GauGAN, Park et al.

slide-11
SLIDE 11

4

GAN Compression Reduces the Computation by 9-21x

MACs

1G 10G 100G 1000G

Original Model GAN Compression

31.7G 4.8G 2.7G 281G 57G 57G

Horse→zebra CycleGAN, Zhu et al. Edge→shoes Pix2pix, Isola et al. Cityscapes GauGAN, Park et al.

slide-12
SLIDE 12

4

GAN Compression Reduces the Computation by 9-21x

MACs

1G 10G 100G 1000G

Original Model GAN Compression

31.7G 4.8G 2.7G 281G 57G 57G

Horse→zebra CycleGAN, Zhu et al. Edge→shoes Pix2pix, Isola et al. Cityscapes GauGAN, Park et al.

21x 12x 9x

slide-13
SLIDE 13

5

Accelerating Horse2zebra by GAN Compression

Original CycleGAN; MACs: 56.8G; FPS: 12.1; FID: 61.5 GAN Compression; MACs: 3.50G (16.2x); FPS: 40.0 (3.3x); FID: 53.6

Measured on NVIDIA Jetson Xavier GPU Lower FID indicates better Performance.

slide-14
SLIDE 14

5

Accelerating Horse2zebra by GAN Compression

Original CycleGAN; MACs: 56.8G; FPS: 12.1; FID: 61.5 GAN Compression; MACs: 3.50G (16.2x); FPS: 40.0 (3.3x); FID: 53.6

Measured on NVIDIA Jetson Xavier GPU Lower FID indicates better Performance.

slide-15
SLIDE 15

Thank you!

A-NVIDIA-SDH Initiative

Hardware, AI and Neural-nets

https://github.com/mit-han-lab

6

Github Page