Variational Autoencoder¶

TODO: intro, link to colab w/ these examples

TODO: math

TODO: diagram

TODO: talk about how this model requires adding additional KL divergence term (for the latent representation posterior vs prior).

import probflow as pf

class VariationalAutoencoder(pf.ContinuousModel):

    def __init__(self, dims)
        self.encoder = pf.DenseRegression(dims, heteroscedastic=True)
        self.decoder = pf.DenseRegression(dims[::-1], heteroscedastic=True)

    def __call__(self, x):
        z = self.encoder(x)
        self.add_kl_loss(z, pf.Normal(0, 1))
        return self.decoder(z.sample())

import probflow as pf
import torch

class VariationalAutoencoder(pf.ContinuousModel):

    def __init__(self, dims)
        self.encoder = pf.DenseRegression(dims, heteroscedastic=True)
        self.decoder = pf.DenseRegression(dims[::-1], heteroscedastic=True)

    def __call__(self, x):
        x = torch.tensor(x)
        z = self.encoder(x)
        self.add_kl_loss(z, pf.Normal(0, 1))
        return self.decoder(z.sample())

Then we can create an instance of the model, defining the dimensionality of each layer of the network:

model = VariationalAutoencoder([7, 128, 64, 32, 3])

TODO: generate some data, and then fit

model.fit(x, x)