Fix Jacobian for dynamic-to-fixed transforms

HISTORY.md

@@ -1,3 +1,7 @@
+# 0.42.5
+
+Fixed `apply_transform_strategy` to accumulate only the target fixed transform's forward log-Jacobian when converting from `DynamicLink`. Previously, the source inverse Jacobian was also included, producing incorrect internal log densities.
+
 # 0.42.4
 
 `arraydist` on a vector of univariate distributions now builds its `Distributions.Product` through the inner constructor instead of `Product(dists)`, which is deprecated. The outer constructor calls `Base.depwarn`, and that walks a backtrace on every call, so models with an `arraydist` likelihood paid it once per evaluation. The return type is unchanged.

Project.toml

@@ -1,6 +1,6 @@
 name = "DynamicPPL"
 uuid = "366bfd00-2699-11ea-058f-f148b4cae6d8"
-version = "0.42.4"
+version = "0.42.5"
 
 [deps]
 ADTypes = "47edcb42-4c32-4615-8424-f2b9edc5f35b"

src/transformed_values.jl

@@ -348,10 +348,7 @@ function apply_transform_strategy(
         fwd_transform = inverse(target.transform)
         transformed_value, logjac = with_logabsdet_jacobian(fwd_transform, raw_value)
         transformed_tv = TransformedValue(transformed_value, target)
-        # TODO: https://github.com/TuringLang/DynamicPPL.jl/issues/1407
-        # Likely should return `logjac` rather than `logjac - inv_logjac`; the sibling
-        # branches all return only the target's forward Jacobian.
-        (raw_value, transformed_tv, logjac - inv_logjac)
+        (raw_value, transformed_tv, logjac)
     else
         error("unknown target transform: $target")
     end

test/transformed_values.jl

@@ -241,6 +241,10 @@ end
 
         # Matching transform should work
         strategy_ok = DynamicPPL.WithTransforms(VarNamedTuple(; x=ft), UnlinkAll())
+        _, _, dynamic_to_fixed_logjac = DynamicPPL.apply_transform_strategy(
+            strategy_ok, TransformedValue(linked_val, DynamicLink()), vn, dist
+        )
+        @test dynamic_to_fixed_logjac ≈ logjac
         new_raw, new_tv, new_logjac = DynamicPPL.apply_transform_strategy(
             strategy_ok, tv, vn, dist
         )