Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Realtime Style Transferについて

48,950 views

Published on

Chainer Meetup #03 LT

Published in: Technology
  • Be the first to comment

Realtime Style Transferについて

  1. 1. Chainer Meetup #03
  2. 2. Yusuke Tomoto / 登本悠介 PG @Rhizomatiks twitter: _mayfa github: yusuketomoto generativeな応用に興味があります。
  3. 3. realtime style transfer
  4. 4. realtime style transfer 1. style transfer 2. realtime style transfer 3. demo
  5. 5. style transfer (neuralstyle, stylenet, neuralart...) image courtesy: https://github.com/jcjohnson/neural-style
  6. 6. “A Neural Algorithm of Artistic Style” Leon A. Gatys, Alexander S. Ecker, Matthias Bethge 2 Sep 2015 VGG-16 style loss content loss content image target style image loss
  7. 7. content loss style loss loss
  8. 8. “A Neural Algorithm of Artistic Style” Leon A. Gatys, Alexander S. Ecker, Matthias Bethge 2 Sep 2015 VGG-16 style loss content loss content image target style image loss backprop!
  9. 9. “A Neural Algorithm of Artistic Style” Leon A. Gatys, Alexander S. Ecker, Matthias Bethge 2 Sep 2015 VGG-16 style loss content loss content image target style image loss backprop!
  10. 10. “A Neural Algorithm of Artistic Style” Leon A. Gatys, Alexander S. Ecker, Matthias Bethge 2 Sep 2015 VGG-16 style loss content loss content image target style image loss backprop!
  11. 11. realtime style transfer
  12. 12. “Perceptual Losses for Real-Time Style Transfer and Super-Resolution” Justin Johnson, Alexandre Alahi, Li Fei-Fei 27 Mar 2016 image transform network generation only one forward pass!
  13. 13. Layer Activation size Input 3 x 256 x 256 32 x 9 x 9 conv, stride 1 32 x 256 x 256 64 x 3 x 3 conv, stride 2 64 x 128 x 128 128 x 3 x 3 conv, stride 2 128 x 64 x 64 Residual block, 128 filters 128 x 64 x 64 Residual block, 128 filters 128 x 64 x 64 Residual block, 128 filters 128 x 64 x 64 Residual block, 128 filters 128 x 64 x 64 Residual block, 128 filters 128 x 64 x 64 64 x 3 x 3 conv, stride 1/2 64 x 128 x 128 32 x 3 x 3 conv, stride 1/2 32 x 256 x 256 3 x 9 x 9 conv, stride 1 3 x 256 x 256 image transform network Architecture
  14. 14. “Perceptual Losses for Real-Time Style Transfer and Super-Resolution” Justin Johnson, Alexandre Alahi, Li Fei-Fei 27 Mar 2016 image transform network VGG16 training content loss style loss total variation loss
  15. 15. “Perceptual Losses for Real-Time Style Transfer and Super-Resolution” Justin Johnson, Alexandre Alahi, Li Fei-Fei 27 Mar 2016 image transform network VGG16 training content loss style loss total variation loss x y^ yc ys
  16. 16. content loss style loss total variation loss
  17. 17. “Perceptual Losses for Real-Time Style Transfer and Super-Resolution” Justin Johnson, Alexandre Alahi, Li Fei-Fei 27 Mar 2016 image transform network VGG16 trainin content loss style loss total variation loss backprop! update model
  18. 18. demo
  19. 19. chainer-fast-neuralstyle https://github.com/yusuketomoto/chainer-fast-neuralstyle ありがとうございました。 *現状ミニバッチ学習(n≧2)にて結果が安定しない不具合あり

×