Inspiration
For a long time, I wanted to implement my own style transfer. I also wanted to learn to use all the cool new stuff inside tensorflow 2. Now both are done :)
What it does
Style transfer is a technology that takes an image as input and apply a style learn by analyzing work of art, for instance Picasso paintings. The content of the image stay the same but the style change.
More technically, this is a generative adversarial network (GAN). A first neural network (the generator) take the image and generate an image with the same content but the style of an artist and a second neural network try to make the difference between this image and a set of images from the artist. Both network train in parallel and at the end, the generator transform an image to another that looks like the images from the artist.
Accomplishments that I'm proud of
Compiling Tensorflow 2 for TensorRT support. It's not very hard but I'm proud to have done it at least once
What I learned
Lot of things about tensorflow 2
What's next for Adaptive-style-transfer-tf2
With this project it is possible to train and export very small models in TFLite, Tensorflow format for embedding device and Android phone. So next step : build an Android app running these models!
Log in or sign up for Devpost to join the conversation.