Fixme in MvNormal

[springcoil]

I came across this - https://github.com/pymc-devs/pymc3/blob/a29f7f1602e239b042128fe0d05a55c482d66162/pymc3/distributions/multivariate.py#L57 does this still need fixed? cc @fonnesbeck @twiecki

A larger follow up question is do we need to run ack or something for all the Fixme's ... I'm trying to not add stuff that needs done to the release though - otherwise it'll take years :)

[twiecki]

Huh, yeah, that looks definitely wrong.

[twiecki]

I'd prefer if we parameterized MvNormal via cov to begin with. Same with Normal which should default to sd instead of tau, there have been two issues where that was the problem. Best to do this as well before a release if it's painful to users in the short-term.

[springcoil]

Just so I understand is this just a case of sensibly picking defaults? So users are exposed to having to learn about this?

[twiecki]

Yes. Most commonly people use standard-deviation and cov to parameterize Normals. However, because of conjugacy, probabilistic programming languages often defaulted to precision instead. Today, conjugacy is not relevant anymore so it's best to make it intuitive.

[springcoil]

• - Fix Normals parameterization
• -Fix MvNormal parameterization

[fonnesbeck]

I agree that parameterizing via precision is a little passé. Happy to switch the default to sd and Sigma, respectively.

[twiecki]

Closed with #1398.