pm.find_MAP won't follow the bounds of custom distribution

Spencer_Hurt · December 6, 2019, 9:39pm

I am defining a custom distribution using the code below:

class Jeffreys(pm.Continuous):
    def __init__(self, lower, upper, **kwargs):
        self.lower = lower
        self.upper = upper
        self.mean = (upper + lower)/2
        self.median = self.mean
        kwargs["testval"] = kwargs.pop("testval", self.mean)
        super(Jeffreys, self).__init__(**kwargs)
    def logp(self, value):
        lower = self.lower
        upper = self.upper
        normalization = 1. / pm.math.log(upper / lower)
        el = tt.log(normalization) - tt.log(value)
        return tt.switch(tt.or_(tt.gt(lower, value), tt.gt(value, upper)), -np.inf, el/2)

When using pm.sample with this distribution in the model, the lower and upper bounds are followed at every point. However, pm.find_MAP consistently goes below the lower bound, returning a logp of -inf. I can solve this problem by using arbitrarily large numbers in place of -np.inf, but I’m curious if there is anything I am doing wrong.

junpenglao · December 9, 2019, 1:58pm

If you have custom distribution that are bounded, you can put a transformation in the __init__ so the bound is taken into account, they best way you could do that is to inherent from BoundedContinuous

github.com

pymc-devs/pymc3/blob/1bf82efc2670eb4b494ab9ea0195165174a58d3c/pymc3/distributions/continuous.py#L46


        super().__init__(transform=transform, *args, **kwargs)




class UnitContinuous(Continuous):
    """Base class for continuous distributions on [0,1]"""


    def __init__(self, transform=transforms.logodds, *args, **kwargs):
        super().__init__(transform=transform, *args, **kwargs)




class BoundedContinuous(Continuous):
    """Base class for bounded continuous distributions"""


    def __init__(self, transform='auto', lower=None, upper=None,
                 *args, **kwargs):


        lower = tt.as_tensor_variable(lower) if lower is not None else None
        upper = tt.as_tensor_variable(upper) if upper is not None else None


        if transform == 'auto':
            if lower is None and upper is None:

Topic		Replies	Views
The Normal distribution with upper and lower bound Questions	2	1729	December 14, 2017
How do i use the Pareto distribution? Questions	2	1199	June 26, 2018
Why cant Uniform or Bounded distributions take upper & lower values as arrays? Questions	2	759	April 28, 2020
Nan while running find_MAP Questions	5	535	December 30, 2019
PyMC3 to PyMC: Alternative to check_parameters and bound? v5	2	505	January 5, 2023

pm.find_MAP won't follow the bounds of custom distribution

Related topics