Hi I’m going to write a distribution which is not in the pymc3’s library. How should I do that? and how should I incorporate that in the codes?? (For example, suppose that Cauchy distribution is not available and I want to write that)

There is a very crude guide in here: Probability Distributions in PyMC3 — PyMC3 3.10.0 documentation You have two options, to create a standard distribution by inheriting from the right subclasses (Discrete, Continuous, PositiveContinuous, etc), or using a pm.Densitydist that takes a callable logp …

Just out of curiosity: what distribution are you planning to create? If commonly used it could end up in the code base, I guess. Also, I’m just about to start writing a few myself (astronomy inspired: power-law distributions, broken power-laws and Schechter functions).

I’m exactly trying to write power-law and Schechter distributions!! :slight_smile: They are so common in astronomy and should be available in pymc3…

If you want to collaborate, hit me up!

Hi, by power-law distribution, do you mean fitting a power law to a power spectra, e.g. a Fourier power spectrum? Also curious as this is what I would like to do.

I keep meaning to add the Generalized Extreme Value distribution too - is this of interest to others? I got as far as realizing that because of theano, it is not as simply as using the SciPy implementation and so it’ll be option 2 for that one.

What we meant here was to implement power-law distribution functions (P(X) ~X**alpha).

Generalized EV might be useful - I’ve written an InverseWeibull for a recent project that I intend to PR once I get a minute!

Writing a new distribution

Questions

fizikdan March 24, 2021, 4:47pm 11

for curious readers!, this is related to the topic

Topic		Replies	Views
Drawing values from custom distribution Questions	8	1165	September 14, 2022
Posterior Predictive Check for Density Dist Questions	4	466	November 17, 2021
Porting PyMC code to run on PyMC3 how to create a stochastic subclass distribution? Questions development	2	920	January 8, 2020
Recommended way to create a new discrete distribution? Questions development	3	1868	February 4, 2020
Using distributions from scipy stats and tensorflow probability in PYMC3 Questions	0	345	October 26, 2021

Writing a new distribution

Related topics