1 / 19

Project on A Neural Algorithm of Artistic Style by Convolutional Neural Networks

Project on A Neural Algorithm of Artistic Style by Convolutional Neural Networks. Yeqi Wang Shuyang Gu. Recap: What we have in the midterm. Read various of paper related to understand exactly how exactly CNN works, what algorithm they use, dig deeper on some details related to the algorithm

dli
Télécharger la présentation

Project on A Neural Algorithm of Artistic Style by Convolutional Neural Networks

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Project on A Neural Algorithm of Artistic Style by Convolutional Neural Networks Yeqi Wang Shuyang Gu

  2. Recap: What we have in the midterm • Read various of paper related to understand exactly how exactly CNN works, what algorithm they use, dig deeper on some details related to the algorithm • I tried to implement all the method followed by the tutorial guide from neural style approach • Using cpu mode(not using cuda), image size: 256*256 • Average iteration time: 5min per 100 iterations

  3. Iteration time : 100 – 400 iterations

  4. What to hope to do: • Still dig deeper on all parameters inside of the CNN, tried to modify some part of Caffe model layout or parameters • Tried to modify the loss function between style and content • Tried to shorten the iteration time based on some modification method: change the input image, change the convolution computation mathods…. • Add some supervised learning feature?

  5. Dig deeper to the CNN model • How loss function works in CNN? • How each layer works during the propagation? • Some tricks we use in the CNN.

  6. Sytle loss: For each layer, we have a “Style loss”, which is more important than content

  7. x • Gram matrix of layer as “feature” -> style loss in layer l • Global feature

  8. Fine tuning CNN layer configuration • In the original design of style of loss layer, it has one single style loss layer in each layer • Tried to add more style loss layer in each layer(after each ReLu layer) • Max pooling -> average pooling

  9. Original

  10. Details: • We may see some difference among details of the generated image after style layer configuration change (even all the other parameters remains the same)

  11. Max pooling -> Avarage pooling

  12. Loss function change

  13. Distribution of W • We tried Gaussian distribution as the simplest one (Assumption: middle layer will have a better style construction)

  14. Gradient decent of weight with Gaussian

  15. What we want from “art style”

  16. Tricks in CNN • Eliminate sizing headaches TIPS/TRICKS • - start with image that has power-of-2 size- for conv layers, use stride 1 filter size 3x3 pad input with aborder of zeros (1 spatially)This makes it so that: [W1,H1,D1] -> [W1,H1,D2] (i.e. spatial sizeexactly preserved)- for pool layers, use pool size 2x2 (more = worse) (slide from Feifei Li & Andrej Karparthy )

  17. Thanks != +

More Related