Deprecated Constant class

[fonnesbeck]

Fixes #2451

Also adds a bunch of newer distributions that are missing from the API docs.

[aseyboldt]

Hm. Having the dirac measure as a subclass of Continuous seems strange from a mathematical point of view. It doesn't have a density.
I can't even think of a use case for this distribution, couldn't we always use a constant (optionally wrapped in a deterministic) instead? Unless someone knows one, I'd rather deprecate it.

[fonnesbeck]

We've had this discussion in the past, but we can bring it up again. However, if someone wants to set a value of a constant to a floating point value they should be able to, and so it makes sense to move it here based on our taxonomy. Alternatively, we could have it not subclass from either Continuous or Discrete.

Moving it to Continuous is the easiest.

[aseyboldt]

Wouldn't that mean that during auto assignment of step methods NUTS will try to sample from that?

[fonnesbeck]

Correct. I was on your side the last time the issue was brought up, but folks argued strongly in favor of keeping it. It apparently does get used, as evidenced by the issue.

[aseyboldt]

Well, then maybe it is best if it inherits from Distribution directly. That way it will at least not mess up step method assignment (I hope?). Although I wouldn't be surprised if we had some code that actually assumes that all distributions are either discrete or continuous.

[fonnesbeck]

We won't know until we try ...

[junpenglao]

couldn't we always use a constant (optionally wrapped in a deterministic) instead?

I agree. Is it possible to just internally take the constant and wrapped in a deterministic?

[twiecki]

Yeah, I don't recall the reasoning last time, but just because people are using it doesn't mean that there isn't a better way which we should support instead (i.e. a Deterministic with a float). I'm in favor of dropping it.

[fonnesbeck]

Ok, I will leave it where it is (Deterministic) and add a deprecation warning.

[fonnesbeck]

OK, after all that, we are just deprecating now.

[springcoil]

Is this ready to merge?

[fonnesbeck]

Please do, unless anyone objects

[springcoil]

Done.

[michaelosthege]

Hmmm. I'm the one who opened #2451 in the first place and now the deprecation of Constant is causing some real trouble for me.
Let me explain:
I'm building a package where a model is independent of the data. For sampling it, one takes a dataset, maps each instance (experiment) in the dataset to prior Distributions and then the package instantiates prior RVs and build the model (computation graph) for the dataset at hand.
To do this, I specify priors as pymc3.Distribution objects (eg. pymc3.Normal.dist("ned", mu=1, sd=2)) outside of any model context.

From that perspective, a Constant is a perfectly fine distribution!
But pymc3.Deterministic is a function, not a distribution. It can not exist outside of a model context. Which means one can no longer describe a constant distribution independent of a model context.

I was able to use pymc3.Deterministic by introducing some if/else in my code, but having the CompundStep skip constant RVs seems a much more elegant solution.