Merge branch 'tor/simple-varinfo-v2' into tor/immutable-varinfo-support

Author: torfjelde

Project.toml

@@ -11,6 +11,7 @@ ChainRulesCore = "d360d2e6-b24c-11e9-a2a3-2a2ae2dbcce4"
 Distributions = "31c24e10-a181-5473-b8eb-7969acd0382f"
 MacroTools = "1914dd2f-81c6-5fcd-8719-6d5c9610ff09"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
+Setfield = "efcf1570-3423-57d1-acb7-fd33fddbac46"
 ZygoteRules = "700de1a5-db45-46bc-99cf-38207098b444"
 
 [compat]

src/DynamicPPL.jl

@@ -34,6 +34,7 @@ export AbstractVarInfo,
     VarInfo,
     UntypedVarInfo,
     TypedVarInfo,
+    SimpleVarInfo,
     push!!,
     empty!!,
     getlogp,
@@ -135,6 +136,7 @@ include("varname.jl")
 include("distribution_wrappers.jl")
 include("contexts.jl")
 include("varinfo.jl")
+include("simple_varinfo.jl")
 include("threadsafe.jl")
 include("context_implementations.jl")
 include("compiler.jl")

src/simple_varinfo.jl

@@ -0,0 +1,278 @@
+using Setfield
+
+"""
+    SimpleVarInfo{NT,T} <: AbstractVarInfo
+
+A simple wrapper of the parameters with a `logp` field for
+accumulation of the logdensity.
+
+Currently only implemented for `NT<:NamedTuple` and `NT<:Dict`.
+
+# Notes
+The major differences between this and `TypedVarInfo` are:
+1. `SimpleVarInfo` does not require linearization.
+2. `SimpleVarInfo` can use more efficient bijectors.
+3. `SimpleVarInfo` is only type-stable if `NT<:NamedTuple` and either
+   a) no indexing is used in tilde-statements, or
+   b) the values have been specified with the corret shapes.
+
+# Examples
+```jldoctest; setup=:(using Distributions)
+julia> using StableRNGs
+
+julia> @model function demo()
+           m ~ Normal()
+           x = Vector{Float64}(undef, 2)
+           for i in eachindex(x)
+               x[i] ~ Normal()
+           end
+           return x
+       end
+demo (generic function with 1 method)
+
+julia> m = demo();
+
+julia> rng = StableRNG(42);
+
+julia> ### Sampling ###
+       ctx = SamplingContext(Random.GLOBAL_RNG, SampleFromPrior(), DefaultContext());
+
+julia> # In the `NamedTuple` version we need to provide the place-holder values for
+       # the variablse which are using "containers", e.g. `Array`.
+       # In this case, this means that we need to specify `x` but not `m`.
+       _, vi = DynamicPPL.evaluate(m, SimpleVarInfo((x = ones(2), )), ctx); vi
+SimpleVarInfo{NamedTuple{(:x, :m), Tuple{Vector{Float64}, Float64}}, Float64}((x = [1.6642061055583879, 1.796319600944139], m = -0.16796295277202952), -5.769094411622931)
+
+julia> # (✓) Vroom, vroom! FAST!!!
+       DynamicPPL.getval(vi, @varname(x[1]))
+1.6642061055583879
+
+julia> # We can also access arbitrary varnames pointing to `x`, e.g.
+       DynamicPPL.getval(vi, @varname(x))
+2-element Vector{Float64}:
+ 1.6642061055583879
+ 1.796319600944139
+
+julia> DynamicPPL.getval(vi, @varname(x[1:2]))
+2-element view(::Vector{Float64}, 1:2) with eltype Float64:
+ 1.6642061055583879
+ 1.796319600944139
+
+julia> # (×) If we don't provide the container...
+       _, vi = DynamicPPL.evaluate(m, SimpleVarInfo(), ctx); vi
+ERROR: type NamedTuple has no field x
+[...]
+
+julia> # If one does not know the varnames, we can use a `Dict` instead.
+       _, vi = DynamicPPL.evaluate(m, SimpleVarInfo{Float64}(Dict()), ctx); vi
+SimpleVarInfo{Dict{Any, Any}, Float64}(Dict{Any, Any}(x[1] => 1.192696983568277, x[2] => 0.4914514300738121, m => 0.25572200616753643), -3.6215377732004237)
+
+julia> # (✓) Sort of fast, but only possible at runtime.
+       DynamicPPL.getval(vi, @varname(x[1]))
+1.192696983568277
+
+julia> # In addtion, we can only access varnames as they appear in the model!
+       DynamicPPL.getval(vi, @varname(x))
+ERROR: KeyError: key x not found
+[...]
+
+julia> julia> DynamicPPL.getval(vi, @varname(x[1:2]))
+ERROR: KeyError: key x[1:2] not found
+[...]
+```
+"""
+struct SimpleVarInfo{NT,T} <: AbstractVarInfo
+    θ::NT
+    logp::T
+end
+
+SimpleVarInfo{T}(θ) where {T<:Real} = SimpleVarInfo{typeof(θ),T}(θ, zero(T))
+SimpleVarInfo(θ) = SimpleVarInfo{eltype(first(θ))}(θ)
+SimpleVarInfo{T}() where {T<:Real} = SimpleVarInfo{T}(NamedTuple())
+SimpleVarInfo() = SimpleVarInfo{Float64}()
+
+# Constructor from `Model`.
+SimpleVarInfo(model::Model, args...) = SimpleVarInfo{Float64}(model, args...)
+function SimpleVarInfo{T}(model::Model, args...) where {T<:Real}
+    _, svi = DynamicPPL.evaluate(model, SimpleVarInfo{T}(), args...)
+    return svi
+end
+
+# Constructor from `VarInfo`.
+function SimpleVarInfo(vi::TypedVarInfo, ::Type{D}=NamedTuple; kwargs...) where {D}
+    return SimpleVarInfo{eltype(getlogp(vi))}(vi, D; kwargs...)
+end
+function SimpleVarInfo{T}(
+    vi::VarInfo{<:NamedTuple{names}}, ::Type{D}
+) where {T<:Real,names,D}
+    values = values_as(vi, D)
+    return SimpleVarInfo{T}(values)
+end
+
+getlogp(vi::SimpleVarInfo) = vi.logp
+setlogp!!(vi::SimpleVarInfo, logp) = SimpleVarInfo(vi.θ, logp)
+acclogp!!(vi::SimpleVarInfo, logp) = SimpleVarInfo(vi.θ, getlogp(vi) + logp)
+
+function setlogp!!(vi::SimpleVarInfo{<:Any,<:Ref}, logp)
+    vi.logp[] = logp
+    return vi
+end
+
+function acclogp!!(vi::SimpleVarInfo{<:Any,<:Ref}, logp)
+    vi.logp[] += logp
+    return vi
+end
+
+function _getvalue(nt::NamedTuple, ::Val{sym}, inds=()) where {sym}
+    # Use `getproperty` instead of `getfield`
+    value = getproperty(nt, sym)
+    # Note that this will return a `view`, even if the resulting value is 0-dim.
+    # This makes it possible to call `setindex!` on the result later to update
+    # in place even in the case where are retrieving a single element, e.g. `x[1]`.
+    return _getindex(value, inds)
+end
+
+# `NamedTuple`
+function getval(vi::SimpleVarInfo{<:NamedTuple}, vn::VarName{sym}) where {sym}
+    return maybe_unwrap_view(_getvalue(vi.θ, Val{sym}(), vn.indexing))
+end
+
+# `Dict`
+function getval(vi::SimpleVarInfo{<:Dict}, vn::VarName)
+    return vi.θ[vn]
+end
+
+# `SimpleVarInfo` doesn't necessarily vectorize, so we can have arrays other than
+# just `Vector`.
+getval(vi::SimpleVarInfo, vns::AbstractArray{<:VarName}) = map(vn -> getval(vi, vn), vns)
+# To disambiguiate.
+getval(vi::SimpleVarInfo, vns::Vector{<:VarName}) = map(vn -> getval(vi, vn), vns)
+
+haskey(vi::SimpleVarInfo, vn) = haskey(vi.θ, getsym(vn))
+
+istrans(::SimpleVarInfo, vn::VarName) = false
+
+getindex(vi::SimpleVarInfo, spl::SampleFromPrior) = vi.θ
+getindex(vi::SimpleVarInfo, spl::SampleFromUniform) = vi.θ
+# TODO: Should we do better?
+getindex(vi::SimpleVarInfo, spl::Sampler) = vi.θ
+getindex(vi::SimpleVarInfo, vn::VarName) = getval(vi, vn)
+getindex(vi::SimpleVarInfo, vns::AbstractArray{<:VarName}) = getval(vi, vns)
+# HACK: Need to disambiguiate.
+getindex(vi::SimpleVarInfo, vns::Vector{<:VarName}) = getval(vi, vns)
+
+# Necessary for `matchingvalue` to work properly.
+function Base.eltype(
+    vi::SimpleVarInfo{<:Any,T}, spl::Union{AbstractSampler,SampleFromPrior}
+) where {T}
+    return T
+end
+
+# `NamedTuple`
+function push!!(
+    vi::SimpleVarInfo{<:NamedTuple},
+    vn::VarName{sym,Tuple{}},
+    value,
+    dist::Distribution,
+    gidset::Set{Selector},
+) where {sym}
+    @set vi.θ = merge(vi.θ, NamedTuple{(sym,)}((value,)))
+end
+function push!!(
+    vi::SimpleVarInfo{<:NamedTuple},
+    vn::VarName{sym},
+    value,
+    dist::Distribution,
+    gidset::Set{Selector},
+) where {sym}
+    # We update in place.
+    # We need a view into the array, hence we call `_getvalue` directly
+    # rather than `getval`.
+    current = _getvalue(vi.θ, Val{sym}(), vn.indexing)
+    current .= value
+    return vi
+end
+
+# `Dict`
+function push!!(
+    vi::SimpleVarInfo{<:Dict}, vn::VarName, r, dist::Distribution, gidset::Set{Selector}
+)
+    vi.θ[vn] = r
+    return vi
+end
+
+# Context implementations
+function tilde_assume!!(context, right, vn, inds, vi::SimpleVarInfo)
+    value, logp, vi_new = tilde_assume(context, right, vn, inds, vi)
+    return value, acclogp!!(vi_new, logp)
+end
+
+function assume(dist::Distribution, vn::VarName, vi::SimpleVarInfo)
+    left = vi[vn]
+    return left, Distributions.loglikelihood(dist, left), vi
+end
+
+function assume(
+    rng::Random.AbstractRNG,
+    sampler::SampleFromPrior,
+    dist::Distribution,
+    vn::VarName,
+    vi::SimpleVarInfo,
+)
+    value = init(rng, dist, sampler)
+    vi = push!!(vi, vn, value, dist, sampler)
+    vi = settrans!!(vi, false, vn)
+    return value, Distributions.loglikelihood(dist, value), vi
+end
+
+# function dot_tilde_assume!!(context, right, left, vn, inds, vi::SimpleVarInfo)
+#     throw(MethodError(dot_tilde_assume!!, (context, right, left, vn, inds, vi)))
+# end
+
+function dot_tilde_assume!!(context, right, left, vn, inds, vi::SimpleVarInfo)
+    value, logp, vi_new = dot_tilde_assume(context, right, left, vn, inds, vi)
+    # Mutation of `value` no longer occurs in main body, so we do it here.
+    left .= value
+    return value, acclogp!!(vi_new, logp)
+end
+
+function dot_assume(
+    dist::MultivariateDistribution,
+    var::AbstractMatrix,
+    vns::AbstractVector{<:VarName},
+    vi::SimpleVarInfo,
+)
+    @assert length(dist) == size(var, 1)
+    # NOTE: We cannot work with `var` here because we might have a model of the form
+    #
+    #     m = Vector{Float64}(undef, n)
+    #     m .~ Normal()
+    #
+    # in which case `var` will have `undef` elements, even if `m` is present in `vi`.
+    value = vi[vns]
+    lp = sum(zip(vns, eachcol(value))) do vn, val
+        return Distributions.logpdf(dist, val)
+    end
+    return value, lp, vi
+end
+
+function dot_assume(
+    dists::Union{Distribution,AbstractArray{<:Distribution}},
+    var::AbstractArray,
+    vns::AbstractArray{<:VarName},
+    vi::SimpleVarInfo{<:NamedTuple},
+)
+    # NOTE: We cannot work with `var` here because we might have a model of the form
+    #
+    #     m = Vector{Float64}(undef, n)
+    #     m .~ Normal()
+    #
+    # in which case `var` will have `undef` elements, even if `m` is present in `vi`.
+    value = vi[vns]
+    lp = sum(Distributions.logpdf.(dists, value))
+    return value, lp, vi
+end
+
+# HACK: Allows us to re-use the impleemntation of `dot_tilde`, etc. for literals.
+increment_num_produce!(::SimpleVarInfo) = nothing
+settrans!!(vi::SimpleVarInfo, trans::Bool, vn::VarName) = vi

src/varinfo.jl

@@ -1398,7 +1398,7 @@ function setval!(
     return setval!(vi, chains.value[sample_idx, :, chain_idx], keys(chains))
 end
 
-function _setval_kernel!(vi::AbstractVarInfo, vn::VarName, values, keys)
+function _setval_kernel!(vi::VarInfo, vn::VarName, values, keys)
     indices = findall(Base.Fix1(subsumes_string, string(vn)), keys)
     if !isempty(indices)
         val = reduce(vcat, values[indices])
@@ -1479,7 +1479,7 @@ function setval_and_resample!(
     return setval_and_resample!(vi, chains.value[sample_idx, :, chain_idx], keys(chains))
 end
 
-function _setval_and_resample_kernel!(vi::AbstractVarInfo, vn::VarName, values, keys)
+function _setval_and_resample_kernel!(vi::VarInfo, vn::VarName, values, keys)
     indices = findall(Base.Fix1(subsumes_string, string(vn)), keys)
     if !isempty(indices)
         val = reduce(vcat, values[indices])
@@ -1493,3 +1493,26 @@ function _setval_and_resample_kernel!(vi::AbstractVarInfo, vn::VarName, values,
 
     return indices
 end
+
+"""
+    values_as(vi::TypedVarInfo, ::Type{NamedTuple})
+    values_as(vi::TypedVarInfo, ::Type{Dict})
+
+Return values in `vi` as the specified type, e.g. `NamedTuple` is returned if
+"""
+function values_as(vi::VarInfo{<:NamedTuple{names}}, ::Type{NamedTuple}) where {names}
+    iter = Iterators.flatten(values_from_metadata(getfield(vi.metadata, n)) for n in names)
+    return NamedTuple(map(p -> Symbol(p.first) => p.second, iter))
+end
+
+function values_as(vi::VarInfo{<:NamedTuple{names}}, ::Type{Dict}) where {names}
+    iter = Iterators.flatten(values_from_metadata(getfield(vi.metadata, n)) for n in names)
+    return Dict(iter)
+end
+
+function values_from_metadata(md::Metadata)
+    return (
+        vn => reconstruct(md.dists[md.idcs[vn]], md.vals[md.ranges[md.idcs[vn]]]) for
+        vn in md.vns
+    )
+end