Labels of means from variational approximations?

[sammosummo]

If I have a model with lots of stochastic RVs, and I approximate the model using variational inference using approx = pm.fit(), approx.mean.eval() returns, I assume, the means of the approximate posterior distributions for all stochastic RVs. But this method doesn’t give the labels, so I’m not sure which value corresponds to which RV. Any suggestions on how to map them?

[junpenglao]

There is a mapping property, see:

Extracting variable means from ADVI fitted model

In a PyMC3 model there is a property to map dict to array and vice versa under model.bijection. A similar property is in the approximation class, for example, you can do: approx = pm.fit() # dict means_dict = approx.bij.rmap(approx.params[0].eval()) # array sample_array = approx.bij.map(approx.sample(1))

    approx = pm.fit()
    # dict
    means_dict = approx.bij.rmap(approx.params[0].eval())
    # array
    sample_array = approx.bij.map(approx.sample(1))

[sammosummo]

Fantastic. Thanks!