KeywordCalls compatibility #93
Conversation
probability => parameterized
|
|
|
|
|
|
I had to make some adjustments in KeywordCalls. Best to use version 1.8 of that, or (until that's registered) this branch: |
| ∫(::typeof(identity), ::Dists.Distribution) = 1.0 | ||
|
|
||
| logdensity(μ::Dists.Distribution, x) = Dists.logpdf(μ,x) | ||
| Dists.logpdf(d::AbstractMeasure, x) = logdensity(d,x) |
There was a problem hiding this comment.
should this mimic distribution's pdf, which is kind of broken? If we would like to represent the actual behaviour
function Dists.logpdf(d::AbstractMeasure, x)
μ = basemeasure(d)
if μ ≪ Lebesgue(...) # depending on size?
logdensity(d,Lebesgue, x)
elseif μ ≪ CountingMeasure(...)
logdensity(d,CountingMeasure, x)
else
error("logpdf undefined, try logdensity")
end
end
There was a problem hiding this comment.
Ahh, good point. I'd like to avoid the conditionals if possible, since they can make it hard to extend. But this is the idea of representative. There are actually some pending design decisions around that concept. This PR review won't be very visible, so maybe I'll start a new issue to discuss this.
There was a problem hiding this comment.
@mschauer I added an issue on this here:
#98
That's maybe not entirely clear - giving a clear description of something as I'm figuring it out is kind of tricky. I think what it comes down to is that we can have
function Dists.logpdf(d::AbstractMeasure, x)
b = fix(basemeasure, d)
return logdensity(d, b)
endThis traverses to a fixpoint twice (once to find the fixpoint, and again to compute the logdensity). I'm sure we can make this more efficient. When we do, we can add a test
d = ...
x = rand(d)
@test Dists.logpdf(d, x) == logdensity(d, fix(basemeasure, d), x)What do you think?
|
@mschauer I'm going to go ahead with this, with the understanding that some of the "representative measure" stuff is still very unstable |
2 participants
Changes
Dependencies
IsPrimitiveandIsRepresentative, using SimpleTraits.jl (seetraits.jl)transforms.jlRenaming
@measureis too vague, changed to@parameterizedElementwiseProductMeasureis a bit much, it's nowPointwiseProductMeasureLikelihoodis nowLogLikelihood, and it's callable as a functionsrc/probabilityis nowsrc/parameterizedNew stuff
@σ_methods, mostly for use with measures on postive realsasparamsworking better@primitiveto declare a measure to have theIsPrimitivetrait (very simple, really just to save end users from having to worry about traits)