More compat updates (#27)

* Make tests pass with StaticArrays

* Fix remaining adjoint tests

* Bump patch

* Bump Distributions compat

Project.toml

@@ -1,7 +1,7 @@
 name = "TemporalGPs"
 uuid = "e155a3c4-0841-43e1-8b83-a0e4f03cc18f"
 authors = ["willtebbutt <wt0881@my.bristol.ac.uk>"]
-version = "0.3.5"
+version = "0.3.6"
 
 [deps]
 BlockArrays = "8e7c35d0-a365-5155-bbbb-fb81a777f24e"
@@ -19,10 +19,10 @@ ZygoteRules = "700de1a5-db45-46bc-99cf-38207098b444"
 [compat]
 BlockArrays = "0.12"
 BlockDiagonals = "0.1"
-Distributions = "0.19, 0.20, 0.21, 0.22, 0.23"
+Distributions = "0.19, 0.20, 0.21, 0.22, 0.23, 0.24"
 FillArrays = "0.7, 0.8"
 Kronecker = "0.4"
-StaticArrays = "= 0.12.1"
+StaticArrays = "0.12.1"
 Stheno = "0.6.6"
 Zygote = "0.5"
 ZygoteRules = "0.2"

src/models/immutable_inference_pullbacks.jl

@@ -99,7 +99,6 @@ for (foo, step_foo, foo_pullback, step_foo_pullback) in [
         @assert length(model) == length(ys)
         T = length(model)
 
-
         # Pre-allocate for filtering distributions. The indexing is slightly different for
         # these than for other quantities. In particular, xs[t] := x_{t-1}.
         xs = Vector{typeof(model.gmm.x0)}(undef, T + 1)

test/models/lgssm.jl

@@ -143,6 +143,8 @@ println("lgssm:")
             if eltype(model) == Float64
 
                 # Verify the gradients w.r.t. sampling from the model.
+                P_U = cholesky(x.P).U
+                P_U = storage.val isa SArrayStorage ? UpperTriangular(SMatrix(P_U.data)) : P_U
                 adjoint_test(
                     (As, as, sqrt_Qs, Hs, hs, m, sqrt_P, sqrt_Σs) -> begin
                         Qs = map(U->UpperTriangular(U)'UpperTriangular(U), sqrt_Qs)
@@ -154,7 +156,7 @@ println("lgssm:")
                         return rand(MersenneTwister(123456), model)
                     end,
                     [randn(rng, Dobs) for _ in 1:N],
-                    As, as, sqrt_Qs, Hs, hs, x.m, cholesky(x.P).U, sqrt_Σs;
+                    As, as, sqrt_Qs, Hs, hs, x.m, P_U, sqrt_Σs;
                     rtol=1e-6, atol=1e-6,
                 )
 
@@ -170,7 +172,7 @@ println("lgssm:")
                         return logpdf(model, y)
                     end,
                     randn(rng),
-                    As, as, sqrt_Qs, Hs, hs, x.m, cholesky(x.P).U, sqrt_Σs, y;
+                    As, as, sqrt_Qs, Hs, hs, x.m, P_U, sqrt_Σs, y;
                     atol=1e-5, rtol=1e-5,
                 )
 
@@ -186,7 +188,7 @@ println("lgssm:")
                         return TemporalGPs.whiten(model, y)
                     end,
                     [randn(rng, Dobs) for _ in 1:N],
-                    As, as, sqrt_Qs, Hs, hs, x.m, cholesky(x.P).U, sqrt_Σs, y;
+                    As, as, sqrt_Qs, Hs, hs, x.m, P_U, sqrt_Σs, y;
                     atol=1e-6, rtol=1e-6,
                 )
 
@@ -202,7 +204,7 @@ println("lgssm:")
                         return TemporalGPs.unwhiten(model, α)
                     end,
                     [randn(rng, Dobs) for _ in 1:N],
-                    As, as, sqrt_Qs, Hs, hs, x.m, cholesky(x.P).U, sqrt_Σs, y;
+                    As, as, sqrt_Qs, Hs, hs, x.m, P_U, sqrt_Σs, y;
                     atol=1e-6, rtol=1e-6,
                 )
             end

test/models/scalar_lgssm.jl

@@ -66,6 +66,8 @@ println("scalar_lgssm:")
         sqrt_Σs = map(Σ->cholesky(Symmetric(Σ)).U, Σs)
 
         # Verify the gradients w.r.t. sampling from the model.
+        P_U = cholesky(x.P).U
+        P_U = storage.val isa SArrayStorage ? UpperTriangular(SMatrix(P_U.data)) : P_U
         adjoint_test(
             (As, as, sqrt_Qs, Hs, hs, m, sqrt_P, sqrt_Σs) -> begin
                 Qs = map(U->UpperTriangular(U)'UpperTriangular(U), sqrt_Qs)
@@ -77,7 +79,7 @@ println("scalar_lgssm:")
                 return rand(MersenneTwister(123456), scalar_model)
             end,
             randn(rng, N),
-            As, as, sqrt_Qs, Hs, hs, x.m, cholesky(x.P).U, sqrt_Σs;
+            As, as, sqrt_Qs, Hs, hs, x.m, P_U, sqrt_Σs;
             rtol=1e-6, atol=1e-6,
         )
 
@@ -93,7 +95,7 @@ println("scalar_lgssm:")
                 return logpdf(scalar_model, y)
             end,
             randn(rng),
-            As, as, sqrt_Qs, Hs, hs, x.m, cholesky(x.P).U, sqrt_Σs, y;
+            As, as, sqrt_Qs, Hs, hs, x.m, P_U, sqrt_Σs, y;
             atol=1e-6, rtol=1e-6,
         )
 
@@ -109,7 +111,7 @@ println("scalar_lgssm:")
                 return TemporalGPs.whiten(scalar_model, y)
             end,
             randn(rng, N),
-            As, as, sqrt_Qs, Hs, hs, x.m, cholesky(x.P).U, sqrt_Σs, y;
+            As, as, sqrt_Qs, Hs, hs, x.m, P_U, sqrt_Σs, y;
             atol=1e-6, rtol=1e-6,
         )
 
@@ -125,7 +127,7 @@ println("scalar_lgssm:")
                 return TemporalGPs.unwhiten(scalar_model, α)
             end,
             randn(rng, N),
-            As, as, sqrt_Qs, Hs, hs, x.m, cholesky(x.P).U, sqrt_Σs, y;
+            As, as, sqrt_Qs, Hs, hs, x.m, P_U, sqrt_Σs, y;
             atol=1e-6, rtol=1e-6,
         )
     end