add VPG #733
add VPG #733
Conversation
| if p.dist <: DiscreteDistribution | ||
| log_prob = s |> A |> logsoftmax | ||
| log_probₐ = log_prob[CartesianIndex.(a, 1:length(a))] | ||
| elseif p.dist <: ContinuousDistribution | ||
| dist = p.dist.(A(s)...) # TODO: this part does not work on GPU. See: https://github.com/JuliaStats/Distributions.jl/issues/1183 . | ||
| log_probₐ = logpdf.(dist, A) | ||
| end |
There was a problem hiding this comment.
i'm not sure this would work. dist maps the output of the network to a distribution, so it's a function type in general. in the discrete case, where dist = Categorical, Categorical is coincidentally the type as well as the constructor, so p.dist <: DiscreteDistribution would match. but in general, dist need not be a Distribution type.
There was a problem hiding this comment.
Yes, here dist is a mapping from parameters... -> distribution. A more conventional way is to add a custom trait to specify the distribution type of the result. But given that most distributions provided in Distributions.jl are both type and function, I think it's safe to assume dist is always a struct type here.
There was a problem hiding this comment.
oh! the example i had in mind was Normal(μ,σ) which has two arguments but i guess that works too with A(s)..., nice
wait JuliaRegistries/General#64390