zhusuan.transform¶
-
planar_normalizing_flow
(samples, log_probs, n_iters)¶ Perform Planar Normalizing Flow along the last axis of inputs.
\[f(z_t) = z_{t-1} + h(z_{t-1} * w_t + b_t) * u_t\]with activation function tanh as well as the invertibility trick from (Danilo 2016).
Parameters: - samples – A N-D (N>=2) float32 Tensor of shape […, d], and planar normalizing flow will be performed along the last axis.
- log_probs – A (N-1)-D float32 Tensor, should be of the same shape as the first N-1 axes of samples.
- n_iters – A int, which represents the number of successive flows.
Returns: A N-D Tensor, the transformed samples.
Returns: A (N-1)-D Tensor, the log probabilities of the transformed samples.