How to query deterministic transformations of StochasticTensor? about zhusuan

[wmyw96]

Suppose you want to perform transform on $w$, you can get the samples and log_probs of $w$ from BayesianNet.query, modify them by $f_{MLP}(z)$ defined by you own, and the use the transformed samples and log_probs in inference algorithm. Moreover, we implement the Normalizing Flow in examples/normalizing_flows, you can refer to it as examples.

[meta-inf]

Thanks. Guess I'm a bit unconscious now ...

[thjashin]

@meta-inf The case is that when you do deterministic transformation, unless it's invertible (one-to-one), there is no easy way to compute the transformed probability. But getting its sample is easy. You could just return it from your model building code.

def build_model():
    with zs.BayesianNet() as model:
        a = zs.Normal('a')
        # deterministic transform 
        b = f(a)
    return model, b

For invertible mapping, we provided some support from the recent literature (normalizing flow: http://proceedings.mlr.press/v37/rezende15.pdf)