Using the compile_logp function to evaluate the log-posterior of a point estimate

[ricardoV94]

@Antonis_Mav

Now, I can calculate the log posterior by setting transform=None. Does this option impact the inference?

Depends on what you want to do with the model. NUTS and find_MAP will struggle with transform=None because of the hard parameter constraints when transformations are not used.

@Antonis_Mav

Also, I can give you some more specifics of what I want to do. My idea is to derive a point estimate (such as the mean of the posterior or MAP) from the posterior trace. This leads to a dictionary with keys being the model.free_RVs.

And you want to evaluate the model logp at this mean? Or you just want the mean? If you want to use it as an input, you will have to transform it. We should probably add an option to disable transforms in compile_logp, to make this easier.

@Antonis_Mav

In case I don’t use transform=None, how would I know the transformed variable names in order to create the point estimate dictionary? Also, would I have to calculate manually their point estimate by transforming the trace accordingly?

You can obtain the value variable that corresponds to each rv by doing model.rvs_to_values[rv]. You can then check the name of this variable.

import pymc as pm

with pm.Model() as m:
    x = pm.Normal("x")
    y = pm.HalfNormal("y")
    z = pm.Normal("z", x, y, observed=0)

value_vars = [m.rvs_to_values[rv] for rv in m.free_RVs]
value_var_names = [v.name for v in value_vars]
value_var_transforms = [getattr(v.tag, "transform", None) for v in value_vars]

print(value_var_names)
print(value_var_transforms)

# ['x', 'y_log__']
# [None, <aeppl.transforms.LogTransform object at 0x7f78de2f01c0>]