Hierachical Temporal Difference Variational Auto-Encoder with jumpy predictions.

Temporal Difference Variational Auto-Encoder, Karol Gregor, George Papamakarios, Frederic Besse, Lars Buesing and Theophane Weber, ICLR 2019, https://openreview.net/forum?id=S1x4ghC9tQ.

First, let’s first go through some definitions which would help understanding what is going on in the following code:

Observation: the observated variable x.

Belief: as the model is feed with a sequence of observations, x_t, the model updates its belief state b_t, through a LSTM network. It is a deterministic function of x_t. We call b_t the belief at time t instead of belief state.

State: the latent hidden state variable z.

__init__(x_dim, b_dim, z_dim, t, d, n_layers=2, n_lstm_layers=1, preproc_dim=None, add_sigmoid=True)[source]¶

Init class.

Parameters

x_dim : int

the dimension of observed data.

b_dim : int

the belief code dimension.

z_dim : int

the dimension of latent space.

t : int

in jumpy state modeling, t1 can be chosen uniformly from the sequence U(1,t).

d : int

in jumpy state modeling, t2 − t1 can be chosen uniformly over some finite range U(1,d).

n_layers : int, default 2

the number of hierachical level in the model.

n_lstm_layers : int, default 1

the number of recurrent layers, eg setting this paramter to 2 would mean stacking two LSTMs together to form a stacked LSTM, with the second LSTM taking in outputs of the first LSTM and computing the final results.

preproc_dim : int, default None

the dimension of preprocessed observations. If not specified no preprocessing is applied.

add_sigmoid : bool, default True

apply sigmoid activation fct to the decoder.

forward(x)[source]¶

The forward method.

reparameterize(q)[source]¶

The reparametrization trick.

rollout(x, t1, t2)[source]¶

Jumpy rollout.

Parameters

x : Tensor

the input sequences.

t1 : int

the time jump number of steps.

t2 : int

the prediction interval t2 - t1.