Neural Style Transfer on AWS

Slide deck presented at the Luxembourg AWS Meetup on 19 Sep 2019

Transcript

1. Neural Style Transfer on AWS Luxembourg AWS Meetup 19th September 2019 Francesco Pochetti – Senior Data Scientist @Mash & AWS ML Hero co-developed with Gabriele Lanaro – ML Engineer @Comtravo
2. Agenda • Demystifying Neural Style Transfer The Deep Learning magic • VisualNeurons.com Style an image. The traditional way. Style a GIF Style an image. The fast way.
3. https://arxiv.org/abs/1508.06576
4. Random Noise Image Content Loss + Style Loss Content Image Style Image Extracted Style Content Reconstruction Neural Style Transfer
5. Minimize MSE Loss Function between two sets of ConvLayer Activations The Original Image is passed through VGG16 The Random Noise Image is passed through VGG16 Pixels’ values of the Noise Image are tuned until the Original Image is reconstructed Content Reconstruction
6. The Original Style Image is passed through VGG16 256 x 60 x 60 Style Reconstruction - Gram Matrix 256 x 3600 3600 x 256 x 256 x 256 Style Image Gram Matrix Random Noise Image Gram Matrix Minimize MSE Loss Function Pixels values of the Noise Image are tuned until the Style of the Original Image is captured i j i x j 256 x 256
8. Style an image. The traditional way
9. Style a GIF
10. Style an image. The fast way https://arxiv.org/abs/1603.08155
11. Content Loss Style Loss Total Loss (to be minimized) Content Style Result (after training) Output (to be optimized) Transform Net Style an image. The fast way Random Noise Input
12. Style an image. The fast way
13. Q&A https://francescopochetti.com/ for detailed write-ups around Style Transfer on AWS