Porting PyMC code to run on PyMC3 how to create a stochastic subclass distribution?

Questions
1

Hi,
I’m trying to get my head round PyMC3 in order to port some existing code written (and working) for PyMC. I need to create a custom distribution that is a stochastic subclass.

The original PyMC code used functions such as pymc.valuewrapper() and pymc.new_dist_class(). Here is a summary of the PyMC code:

def a_logp(...):
   ...
   return something 

def a_random(...):
    ...
    return something

dtype = np.float64
docstr = "mydist"
parent_names = ["a", "b"]
parents_default = {"a": 0, "b": 1, "value": 0.5}
wrapped_logp_partial_gradients = {}
logp = pymc.valuewrapper(a_logp)
mv = False

UniformNonzero = pymc.new_dist_class(
    dtype,
    node_name,
    parent_names,
    parents_default,
    docstr,
    logp,
    a_random,
    mv,
    wrapped_logp_partial_gradients,
)

My questions are, does PyMC3 support equivalent functionality to do this? I see PyMC3 supports a function called DensityDist() for creating custom distributions using logp() and random() functions. But how do I define the parent information etc? Or is it I have to dig deeper, e.g. into Theano? I would be very grateful if someone could just point in the right direction.

2

Yes you will use DensityDist for this. The conditional (i.e., parent information) could be coded in the logp function directly, or feeded using a dict in observed={…}

You can have a look at How to set up a custom likelihood function for two variables for more information.

3

Many thanks!