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

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 University 1

Conditional GANs are Computationally-Intensive. 2

Conditional GANs are Computationally-Intensive. Pix2pix, CVPR’17 2

Conditional GANs are Computationally-Intensive. MobileNet ResNet-50 CycleGAN GauGAN Pix2pix, CVPR’17 MACs 2

Conditional GANs are Computationally-Intensive. 500x MobileNet 0.5G ResNet-50 4.0G CycleGAN 57G GauGAN 281G Pix2pix, CVPR’17 MACs GANs requires 10 × -500 × more computations than image recognition models! 2

General-purpose Compression Framework—GAN Compression Distillation + Channel Pruning + NAS 3

GAN Compression Reduces the Computation by 9-21x Horse → zebra Edge → shoes Cityscapes CycleGAN, Zhu et al. Pix2pix, Isola et al. GauGAN, Park et al. 4

GAN Compression Reduces the Computation by 9-21x 1000G Original Model GAN Compression 100G MACs 10G 1G Horse → zebra Edge → shoes Cityscapes CycleGAN, Zhu et al. Pix2pix, Isola et al. GauGAN, Park et al. 4

GAN Compression Reduces the Computation by 9-21x 1000G Original Model GAN Compression 100G MACs 57G 10G 2.7G 1G Horse → zebra Edge → shoes Cityscapes CycleGAN, Zhu et al. Pix2pix, Isola et al. GauGAN, Park et al. 4

GAN Compression Reduces the Computation by 9-21x 1000G Original Model GAN Compression 100G MACs 57G 57G 10G 4.8G 2.7G 1G Horse → zebra Edge → shoes Cityscapes CycleGAN, Zhu et al. Pix2pix, Isola et al. GauGAN, Park et al. 4

GAN Compression Reduces the Computation by 9-21x 1000G Original Model GAN Compression 281G 100G MACs 57G 57G 31.7G 10G 4.8G 2.7G 1G Horse → zebra Edge → shoes Cityscapes CycleGAN, Zhu et al. Pix2pix, Isola et al. GauGAN, Park et al. 4

GAN Compression Reduces the Computation by 9-21x 1000G Original Model GAN Compression 281G 9x 100G MACs 57G 57G 31.7G 12x 21x 10G 4.8G 2.7G 1G Horse → zebra Edge → shoes Cityscapes CycleGAN, Zhu et al. Pix2pix, Isola et al. GauGAN, Park et al. 4

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. 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. 5

Thank you! Github Page H ardware, A I and N eural-nets https://github.com/mit-han-lab A-NVIDIA-SDH Initiative 6