Handle Real and Any better #96
Conversation
|
Tests are failing because of an issue with the latest ReverseDiff version which has a lot of the functionality in DistributionsAD. This created some conflict. Should be alright after #97 goes in though. |
| @@ -77,7 +77,7 @@ function Base.rand(rng::AbstractRNG, d::MvDiscreteNonParametric{T,P}) where {T,P | |||
| p = probs(d) | |||
| n, k = size(p) | |||
| map(1:k) do j | |||
| draw = rand(rng, P) | |||
| draw = rand(rng, (P === Real ? Float64 : P)) | |||
There was a problem hiding this comment.
That seems like a quite crude heuristic. Ideally we would want P to be
mapreduce(promote_type, 1:n) do i
typeof(p[i, j])
endassuming that all provided probabilities can be promoted to a common non-abstract type in each column (but possibly not the same type for all columns) - I guess that should usually be satisfied?
BTW copying the logic from Distributions.DiscreteNonParametric here unfortunately introduces also the bugs present in the implementation in Distributions (see e.g. JuliaStats/Distributions.jl#1128), which might not be fixed when the upstream bugs are fixed if people don't keep track of Distributions closely.
This PR allows the handling of arrays of Any and Real better by relaxing some type constraints.