Teaching a neural network to draw

Starring Pytorch, LSTMs, GMMs, and negative log likelihood maximization. See the companion blog post.

See also:

• the reference Google Brain paper
• the Jupyter notebook

Data and models

The datasets come from this repository (simplified binary format). They are to be placed in data/.

Trained models are available in models/. They are single-layer Pytorch LSTMs outputting parameters for a 20-components GMM, with a hidden state size of either 128 or 512.

Setup

The code runs with Python >= 3.6 and Pytorch >= 1.0. Other requirements are listed in requirements.txt.