Random in Categorical returns wrong shape for 2D parameter

[junpenglao]

Running:

p = np.random.rand(10, 5)
pm.Categorical.dist(p=p).random()

gives a (10, 5) array instead of (10, ).

cc @ColCarroll

[junpenglao]

Currently there are 2 similar copies of Categorical Random generation in the code:
https://github.com/pymc-devs/pymc3/blob/ea5cbf3e838973b4f992674d0a8a6bd9a443b229/pymc3/distributions/mixture.py#L150-L158

https://github.com/pymc-devs/pymc3/blob/ea5cbf3e838973b4f992674d0a8a6bd9a443b229/pymc3/distributions/discrete.py#L713-L720

We should unify them into one function.

[canyon289]

@junpenglao I would be happy to help here if I can understand a couple of things.

What is the expected output for the function when an (x,y) input is passed as an argument?
Is the Categorical Random in discrete.py or mixture.py more preferred?

Let me know if help here is desired!

[junpenglao]

What is the expected output for the function when an (x,y) input is passed as an argument?

(x, ) or (x, 1), as the last dimension of the input p (with shape (x, y)) is expected to be equal to the number of categories. Currently Categorical Random in discrete.py returns (x, ) and mixture.py returns (x, 1) - it is easier to return (x, 1) in a mixture for indexing so I suggest this to be the default.

[canyon289]

Got it, If you're open to it I would like to try my hand at this one!

Is there any urgency on when this needs to be completed?

I'm still wrapping my head around the specifics and may need to get a couple of reviews before it's nailed down.

[junpenglao]

Go ahead! Let me know if you have any questions ;-)