MarioBonse / Sketch-rnn

Implementation of the model "sketch-RNN" by google for generating sketches with a variational auto encoder

Sketch-rnn

On this project I implemented, with tensorflow 2.0 and keras, Sketch-rnn, a Variational Autoencoder for generating sketches. I followed the original paper.

Architecture

The architecture is the following.

Colab notebook

I have created also a colab notebook where everyone can train the model and check the results.
I have trained the model with two different sketches:

• Carrot
• Cat

Some sketch generated!

Carrot where the hidden variable is sampled from the IID gaussian:

Carrot where the latent variable is encoded from a sketch

Cat where the hidden variable is sampled from the IID gaussian:

Cat where the latent variable is encoded from a sketch

Example of skatches from the Quick, draw! dataset:

There is also a Jupyter notebook for testing the sampling. We can sample from a hidden variable produced by the encoder or form a sample of a IID gaussian. The weights of the trained model are in the model directory.