Fig. 4.
Excerpt of Pyro code that implements a semi-supervised variant of a variational autoencoder, with graphical representation on the right. The latent variable z captures aspects of the face (e.g., shape) that are emotion-irrelevant, while θ parameterizes the distribution Pθ(face|emotion, z). Here, θ are weights in a neural network within the Decoder () function (l. 7). This code builds off Fig. 3 by generating an emotion conditioned on the observed outcome and emotion ratings (l. 3-4), sampling z from its priors (l. 5), and generating a face conditioned on the observed data (l. 6-8).