Finishing Course 4 of Andrew Ng’s deeplearning.ai @Coursera was really fun. Especially Neural Style Transfer (NST). With this technique the style of a famous painting can be transferred to a new image. Here are some of the examples from the course:
As my style-image I decided to go for for Munch’s “The Scream”. I transferred the style of the painting to my content-image, a picture of a person that looked quite similar to Munch’s protagonist. ^^
The Convolutional Neural Network was pre-trained on VGG-19. A new, noisy image was initialized and generated and then the style was transferred gradually. The hyperparameters of the deeplearning.ai model were set to 200 iterations with an Adam optimizer and a (relatively aggressive) learning rate of 2.0. The output was pretty satisfying. And hilarious. 😀
It was recommended to decrease the learning rate and increase the number of iterations, to get even better results. The original NST paper comes from Gatys et al. (2015) and is a great read, if you want to know more details about the Deep Learning architecture.