Fix dimensionality issues of ADVI

src/variational/advi.jl

@@ -1,45 +1,17 @@
-# TODO(torfjelde): Find a better solution.
-struct Vec{N,B} <: Bijectors.Bijector
-    b::B
-    size::NTuple{N, Int}
-end
-
-Bijectors.inverse(f::Vec) = Vec(Bijectors.inverse(f.b), f.size)
-
-Bijectors.output_length(f::Vec, sz) = Bijectors.output_length(f.b, sz)
-Bijectors.output_length(f::Vec, n::Int) = Bijectors.output_length(f.b, n)
-
-function Bijectors.with_logabsdet_jacobian(f::Vec, x)
-    return Bijectors.transform(f, x), Bijectors.logabsdetjac(f, x)
-end
-
-function Bijectors.transform(f::Vec, x::AbstractVector)
-    # Reshape into shape compatible with wrapped bijector and then `vec` again.
-    return vec(f.b(reshape(x, f.size)))
-end
-
-function Bijectors.transform(f::Vec{N,<:Bijectors.Inverse}, x::AbstractVector) where N
-    # Reshape into shape compatible with original (forward) bijector and then `vec` again.
-    return vec(f.b(reshape(x, Bijectors.output_length(f.b.orig, prod(f.size)))))
-end
-
-function Bijectors.transform(f::Vec, x::AbstractMatrix)
-    # At the moment we do batching for higher-than-1-dim spaces by simply using
-    # lists of inputs rather than `AbstractArray` with `N + 1` dimension.
-    cols = Iterators.Stateful(eachcol(x))
-    # Make `init` a matrix to ensure type-stability
-    init = reshape(f(first(cols)), :, 1)
-    return mapreduce(f, hcat, cols; init = init)
-end
-
-function Bijectors.logabsdetjac(f::Vec, x::AbstractVector)
-    return Bijectors.logabsdetjac(f.b, reshape(x, f.size))
-end
+# TODO: Move to Bijectors.jl if we find further use for this.
+"""
+    wrap_in_vec_reshape(f, in_size)
 
-function Bijectors.logabsdetjac(f::Vec, x::AbstractMatrix)
-    return map(eachcol(x)) do x_
-        Bijectors.logabsdetjac(f, x_)
-    end
+Wraps a bijector `f` such that it operates on vectors of length `prod(in_size)` and produces
+a vector of length `prod(Bijectors.output(f, in_size))`.
+"""
+function wrap_in_vec_reshape(f, in_size)
+    vec_in_length = prod(in_size)
+    reshape_inner = Bijectors.Reshape((vec_in_length,), in_size)
+    out_size = Bijectors.output_size(f, in_size)
+    vec_out_length = prod(out_size)
+    reshape_outer = Bijectors.Reshape(out_size, (vec_out_length,))
+    return reshape_outer ∘ f ∘ reshape_inner
 end
 
 
@@ -83,7 +55,7 @@ function Bijectors.bijector(
         if d isa Distributions.UnivariateDistribution
             b
         else
-            Vec(b, size(d))
+            wrap_in_vec_reshape(b, size(d))
         end
     end
 
@@ -106,7 +78,10 @@ meanfield(model::DynamicPPL.Model) = meanfield(Random.default_rng(), model)
 function meanfield(rng::Random.AbstractRNG, model::DynamicPPL.Model)
     # Setup.
     varinfo = DynamicPPL.VarInfo(model)
-    num_params = length(varinfo[DynamicPPL.SampleFromPrior()])
+    # Use linked `varinfo` to determine the correct number of parameters.
+    # TODO: Replace with `length` once this is implemented for `VarInfo`.
+    varinfo_linked = DynamicPPL.link(varinfo, model)
+    num_params = length(varinfo_linked[:])
 
     # initial params
     μ = randn(rng, num_params)
@@ -134,7 +109,10 @@ function AdvancedVI.update(
     td::Bijectors.TransformedDistribution{<:DistributionsAD.TuringDiagMvNormal},
     θ::AbstractArray,
 )
-    μ, ω = θ[1:length(td)], θ[length(td) + 1:end]
+    # `length(td.dist) != length(td)` if `td.transform` changes the dimensionality,
+    # so we need to use the length of the underlying distribution `td.dist` here.
+    # TODO: Check if we can get away with `view` instead of `getindex` for all AD backends.
+    μ, ω = θ[begin:(begin + length(td.dist) - 1)], θ[(begin + length(td.dist)):end]
     return AdvancedVI.update(td, μ, StatsFuns.softplus.(ω))
 end
 

test/variational/advi.jl

@@ -64,5 +64,10 @@
         x0_inv = inverse(b)(z0)
         @test size(x0_inv) == size(x0)
         @test all(x0 .≈ x0_inv)
+
+        # And regression for https://github.com/TuringLang/Turing.jl/issues/2160.
+        q = vi(m, ADVI(10, 1000))
+        x = rand(q, 1000)
+        @test mean(eachcol(x)) ≈ [0.5, 0.5] atol=0.1
     end
 end