eduzhai > Applied Sciences > Engineering >

GAN Slimming All-in-One GAN Compression by A Unified Optimization Framework

  • king
  • (0) Download
  • 20210507
  • Save

... pages left unread,continue reading

Document pages: 19 pages

Abstract: Generative adversarial networks (GANs) have gained increasing popularity invarious computer vision applications, and recently start to be deployed toresource-constrained mobile devices. Similar to other deep models,state-of-the-art GANs suffer from high parameter complexities. That hasrecently motivated the exploration of compressing GANs (usually generators).Compared to the vast literature and prevailing success in compressing deepclassifiers, the study of GAN compression remains in its infancy, so farleveraging individual compression techniques instead of more sophisticatedcombinations. We observe that due to the notorious instability of trainingGANs, heuristically stacking different compression techniques will result inunsatisfactory results. To this end, we propose the first unified optimizationframework combining multiple compression means for GAN compression, dubbed GANSlimming (GS). GS seamlessly integrates three mainstream compressiontechniques: model distillation, channel pruning and quantization, together withthe GAN minimax objective, into one unified optimization form, that can beefficiently optimized from end to end. Without bells and whistles, GS largelyoutperforms existing options in compressing image-to-image translation GANs.Specifically, we apply GS to compress CartoonGAN, a state-of-the-art styletransfer network, by up to 47 times, with minimal visual quality degradation.Codes and pre-trained models can be found atthis https URL.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×