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.
Pixel Recurrent Neural Network
PR12와 함께 이해하는
Jaejun Yoo
Ph.D. Candidate @KAIST
PR12
16TH July, 2017
Pixel Recurrent Neural Network
PR12와 함께 이해하는
Jaejun Yoo
Ph.D. Candidate @KAIST
PR12
16TH July, 2017
Pixel Recurrent Neural Network
PR12와 함께 이해하는
Jaejun Yoo
Ph.D. Candidate @KAIST
PR12
16TH July, 2017
GENERATIVE MODEL!
NIPS 2016 Tutorial: GANs
MASK
MASK
Channel Masks
• Sequential order: R -> G -> B
• Used in input-to-state convolutions
• Two types of masks:
• Channels ...
Receptive Fields
Architecture
Architecture
Row LSTM
Row LSTM : Receptive Field
Architecture
Diagonal LSTM
Diagonal BiLSTM : Receptive Field
EXPERIMENT AND RESULTS
EXPERIMENT AND RESULTS
• Dataset: MNIST, CIFAR-10, and ImageNet
• Method: log-likelihood
Prerequisite: KL Divergence
WIKI: Information theory
Prerequisite: MLE & KL Divergence
EXPERIMENT AND RESULTS
EXPERIMENT AND RESULTS
“Lower is better”
EXPERIMENT AND RESULTS
EXPERIMENT AND RESULTS
EXPERIMENT AND RESULTS
CIFAR-10 (32x32) ImageNet (32x32)
• Deep neural network that sequentially predicts the pixels in an image
along the two spatial dimensions
• Suggests three ...
Reference
• Slides: Hugo Larochelle, Google Brain: Autoregressive Generative Models
with Deep Learning
• Slides: http://sl...
GRAN
[PR12] PixelRNN- Jaejun Yoo
[PR12] PixelRNN- Jaejun Yoo
[PR12] PixelRNN- Jaejun Yoo
[PR12] PixelRNN- Jaejun Yoo
[PR12] PixelRNN- Jaejun Yoo
[PR12] PixelRNN- Jaejun Yoo
[PR12] PixelRNN- Jaejun Yoo
[PR12] PixelRNN- Jaejun Yoo
[PR12] PixelRNN- Jaejun Yoo
[PR12] PixelRNN- Jaejun Yoo
[PR12] PixelRNN- Jaejun Yoo
Upcoming SlideShare
Loading in …5
×

[PR12] PixelRNN- Jaejun Yoo

1,006 views

Published on

Introduction to pixel recurrent neural network (PixelRNN) and PixelCNN (Korean)
Video: https://youtu.be/BvcwEz4VPIQ

Published in: Technology
  • Be the first to comment

[PR12] PixelRNN- Jaejun Yoo

  1. 1. Pixel Recurrent Neural Network PR12와 함께 이해하는 Jaejun Yoo Ph.D. Candidate @KAIST PR12 16TH July, 2017
  2. 2. Pixel Recurrent Neural Network PR12와 함께 이해하는 Jaejun Yoo Ph.D. Candidate @KAIST PR12 16TH July, 2017
  3. 3. Pixel Recurrent Neural Network PR12와 함께 이해하는 Jaejun Yoo Ph.D. Candidate @KAIST PR12 16TH July, 2017
  4. 4. GENERATIVE MODEL! NIPS 2016 Tutorial: GANs
  5. 5. MASK
  6. 6. MASK Channel Masks • Sequential order: R -> G -> B • Used in input-to-state convolutions • Two types of masks: • Channels are connected to themselves • Used in all other subsequent layers • Channels are not connected to themselves • Only used in the first layer
  7. 7. Receptive Fields
  8. 8. Architecture
  9. 9. Architecture
  10. 10. Row LSTM
  11. 11. Row LSTM : Receptive Field
  12. 12. Architecture
  13. 13. Diagonal LSTM
  14. 14. Diagonal BiLSTM : Receptive Field
  15. 15. EXPERIMENT AND RESULTS
  16. 16. EXPERIMENT AND RESULTS • Dataset: MNIST, CIFAR-10, and ImageNet • Method: log-likelihood
  17. 17. Prerequisite: KL Divergence WIKI: Information theory
  18. 18. Prerequisite: MLE & KL Divergence
  19. 19. EXPERIMENT AND RESULTS
  20. 20. EXPERIMENT AND RESULTS “Lower is better”
  21. 21. EXPERIMENT AND RESULTS
  22. 22. EXPERIMENT AND RESULTS
  23. 23. EXPERIMENT AND RESULTS CIFAR-10 (32x32) ImageNet (32x32)
  24. 24. • Deep neural network that sequentially predicts the pixels in an image along the two spatial dimensions • Suggests three novel architectures to achieve this goal • PixelRNN (Row LSTM, Diagonal BiLSTM), PixelCNN (All Convolutional Net) • Combined with efficient convolution preprocessing, both PixelCNN and PixelRNN use this “product of conditionals” approach to great effect. • The model provides a tractable P(x) with the best log-likelihood scores in its family. • Image generation of good quality and diversity SUMMARY
  25. 25. Reference • Slides: Hugo Larochelle, Google Brain: Autoregressive Generative Models with Deep Learning • Slides: http://slazebni.cs.illinois.edu/spring17/lec13_advanced.pdf • Slides: https://www.slideshare.net/neouyghur/pixel-recurrent-neural- networks-73970786 • Code: https://github.com/igul222/pixel_rnn/blob/master/pixel_rnn.py • Related Paper “Generating images with recurrent adversarial networks” • Related Paper “WaveNet”
  26. 26. GRAN

×