Creating Distributions with CustomDist

[jessegrabowski]

Notebook proposal

Title: Creating Distributions with CustomDist

Why should this notebook be added to pymc-examples?

This comes up a lot on the discourse, most recently here. New users struggle with the API since it's a bit closer to pytensor than the rest of PyMC. There's also some intricacy when one doesn't go the dist kwarg route. For example, you need to know to actually write in numpy/scipy for the random method, but you cannot do that anywhere else in PyMC!

Suggested categories:

• Level: Intermediate
• Diataxis type: How-to

Related notebooks

Closest existing notebook is the black-box likelihood, but that's even more advanced than what I'm proposing.

References

CustomDist API docs are already quite nice. I just envision something a bit more hand-holding with more pictures.

[ricardoV94]

I would suggest not emphasizing the random method at all, since the dist can be used for the random part as well

[jessegrabowski]

I agree the majority of the notebook should show how to work though the dist argument, but there should at least be a note at the bottom about the other way.

[jonsedar]

Interesting, I bumped into stuff like this when wanting to create zero-augmented likelihoods (e.g. a bimodal target variable that's either zero or log-normal). I can't promise to spend the time, but if noone else is doing this I'll consider taking it on :)