LogDensityFunction with dimension unpreserving transformations #1149
Description
Hi, I noted a problem with LogDensityFunction. In principle, I believe LogDensityFunction is supposed to be used in both constrained and unconstrained space, depending on the configuration conveyed by vi. The problem is that certain transformations are volume-preserving but dimension-non-preserving. For instance, the transformation associated with a Dirichlet maps between a LogDensityFunction with a vi with transformed turned on. Here is a basic example:
using DynamicPPL
using ReverseDiff
using Distributions
using LogDensityProblems
using ADTypes, AutoReverseDiff
DynamicPPL.@model function dirichlet()
x ~ Dirichlet([1.0, 1.0])
return x
end
m = dirichlet()
vi = DynamicPPL.ldf_default_varinfo(m, DynamicPPL.getlogjoint_internal)
vi = DynamicPPL.set_transformed!!(vi, true)
prob = DynamicPPL.LogDensityFunction(
m, DynamicPPL.getlogjoint_internal, vi; adtype=AutoReverseDiff()
)The correct dimensionality of prob is now 1, not 2.
However, the following assert fails:
@assert LogDensityProblems.dimension(prob) == 1The following density evaluation also fails:
LogDensityProblems.logdensity(prob, [0.0,])This returns the following:
ERROR: BoundsError: attempt to access 1-element Vector{Float64} at index [1:2]
Stacktrace:
[1] throw_boundserror(A::Vector{Float64}, I::Tuple{UnitRange{Int64}})
@ Base ./essentials.jl:14
[2] checkbounds
@ ./abstractarray.jl:699 [inlined]
[3] getindex
@ ./array.jl:936 [inlined]
[4] macro expansion
@ ~/.julia/packages/DynamicPPL/Ut5Ls/src/varinfo.jl:0 [inlined]
[5] unflatten_metadata(metadata::@NamedTuple{x::DynamicPPL.Metadata{…}}, x::Vector{Float64})
@ DynamicPPL ~/.julia/packages/DynamicPPL/Ut5Ls/src/varinfo.jl:389
[6] unflatten
@ ~/.julia/packages/DynamicPPL/Ut5Ls/src/varinfo.jl:369 [inlined]
[7] logdensity_at
@ ~/.julia/packages/DynamicPPL/Ut5Ls/src/logdensityfunction.jl:242 [inlined]
[8] logdensity(f::LogDensityFunction{…}, x::Vector{…})
@ DynamicPPL ~/.julia/packages/DynamicPPL/Ut5Ls/src/logdensityfunction.jl:279
[9] top-level scope
@ REPL[45]:1
Some type information was truncated. Use `show(err)` to see complete types.More importantly, the following also fails:
LogDensityProblems.logdensity(prob, [0.0, 1.0])by returning the following:
ERROR: DimensionMismatch: inconsistent array dimensions
Stacktrace:
[1] logpdf
@ ~/.julia/packages/Distributions/psM3H/src/common.jl:266 [inlined]
[2] accumulate_assume!!
@ ~/.julia/packages/DynamicPPL/Ut5Ls/src/default_accumulators.jl:95 [inlined]
[3] #64
@ ~/.julia/packages/DynamicPPL/Ut5Ls/src/abstract_varinfo.jl:311 [inlined]
[4] map
@ ./tuple.jl:357 [inlined]
[5] map
@ ./namedtuple.jl:266 [inlined]
[6] map
@ ~/.julia/packages/DynamicPPL/Ut5Ls/src/accumulators.jl:204 [inlined]
[7] map_accumulators!!
@ ~/.julia/packages/DynamicPPL/Ut5Ls/src/abstract_varinfo.jl:330 [inlined]
[8] accumulate_assume!!
@ ~/.julia/packages/DynamicPPL/Ut5Ls/src/abstract_varinfo.jl:311 [inlined]
[9] tilde_assume!!(::DefaultContext, right::Dirichlet{…}, vn::VarName{…}, vi::VarInfo{…})
@ DynamicPPL ~/.julia/packages/DynamicPPL/Ut5Ls/src/contexts/default.jl:36
[10] dirichlet(__model__::Model{…}, __varinfo__::VarInfo{…})
@ Main ./REPL[40]:2
[11] _evaluate!!
@ ~/.julia/packages/DynamicPPL/Ut5Ls/src/model.jl:974 [inlined]
[12] evaluate_threadunsafe!!
@ ~/.julia/packages/DynamicPPL/Ut5Ls/src/model.jl:940 [inlined]
[13] evaluate!!(model::Model{…}, varinfo::VarInfo{…})
@ DynamicPPL ~/.julia/packages/DynamicPPL/Ut5Ls/src/model.jl:925
[14] logdensity_at
@ ~/.julia/packages/DynamicPPL/Ut5Ls/src/logdensityfunction.jl:243 [inlined]
[15] logdensity(f::LogDensityFunction{…}, x::Vector{…})
@ DynamicPPL ~/.julia/packages/DynamicPPL/Ut5Ls/src/logdensityfunction.jl:279
[16] top-level scope
@ REPL[49]:1
Some type information was truncated. Use `show(err)` to see complete types.