Use new struct transforms from TV 0.8.21 (#21)

* Update for TransformVars after adding struct transform

* Add kwarg constructor for RpFormFit

* Update compat entry for TransformVariables

* Clear out ConstWrapTV, which relied on internals of TV anyway

* Update docs

* Test robustness; use up-to-date JET kwargs

Author: Ickaser

Project.toml

@@ -43,7 +43,7 @@ Reexport = "1"
 Roots = "2"
 SavitzkyGolay = "0.9.1"
 SpecialFunctions = "1, 2"
-TransformVariables = "0.8.19"
+TransformVariables = "0.8.21"
 TypedTables = "1.4"
 Unitful = "1.20"
 julia = "1.10"

docs/example/fitting_mannitol.jl

@@ -163,16 +163,16 @@ savefig("modelpre.svg"); #md #hide
 # Optimization algorithms are happiest when they can run across all real numbers.
 # So we use TransformVariables.jl to map all reals to positive values of our parameters, with sensible scales.
 # The `TVExp` transform maps all real numbers to positive values, and the `TVScale` transform scales the value to a more reasonable range.
-# The transform `ConstWrapTV` is defined in LyoPronto, and makes a constant callable function from a value.
 
-# Kshf needs to be callable.
-# Rp needs to be a callable, and the `RpFormFit` struct does that; by passing the new values 
-# with Rp as a NamedTuple, the constructor for `ParamObjPikal` will unpack it.
+# Kshf needs to be callable, so we wrap it in ConstPhysProp.
+# Rp needs to be a callable, and the `RpFormFit` struct with fields R0, A1, and A2 does that. 
+# This `as` function uses TransformVariables to create a transform that maps from 4 real 
+# numbers to callable Kshf and Rp, with appropriate scaling.
 
-trans_KRp = as((Kshf = ConstWrapTV() ∘ TVScale(Kshf(0)) ∘ TVExp(),
-                Rp=as((R0 = TVScale(R0) ∘ TVExp(),
+trans_KRp = as((Kshf = as(ConstPhysProp, (TVScale(Kshf(0)) ∘ TVExp(),)),
+                Rp=as(RpFormFit, as((R0 = TVScale(R0) ∘ TVExp(),
                 A1 = TVScale(A1) ∘ TVExp(),
-                A2 = TVScale(A2) ∘ TVExp(),))))
+                A2 = TVScale(A2) ∘ TVExp(),)))))
 ## Or, using a convenience function for the same,
 trans_KRp = KRp_transform_basic(Kshf(0), R0, A1, A2)
 trans_Rp = Rp_transform_basic(R0, A1, A2)

src/LyoPronto.jl

@@ -58,7 +58,6 @@ export gen_sol_pd, obj_pd, gen_nsol_pd, objn_pd
 export KRp_transform_basic, K_transform_basic, Rp_transform_basic, KBB_transform_basic
 export KBB_transform_bounded
 export obj_expT, err_expT, err_expT!, num_errs, nls_pd, nls_pd!
-export ConstWrapTV
 # plotting tools (mostly already done by macros)
 export qrf_integrate
 # End of primary drying

src/paramfits.jl

@@ -14,12 +14,12 @@ end
 Construct a typical transform for fitting Rp.
 """
 function Rp_transform_basic(R0g, A1g, A2g)
-    tr = as((
-        Rp = as((
+    tr = as((;
+        Rp = as(RpFormFit, as((;
             R0 = TVScale(R0g) ∘ TVExp(),
             A1 = TVScale(A1g) ∘ TVExp(),
             A2 = TVScale(A2g) ∘ TVExp(),
-            )),
+            )),)
         ))
     return tr
 end
@@ -29,7 +29,7 @@ end
 Construct a typical transform for fitting Kshf (a.k.a. Kv).
 """
 function K_transform_basic(Kshfg)
-    tr = as((Kshf = ConstWrapTV() ∘ TVScale(Kshfg) ∘ TVExp(),))
+    tr = as((;Kshf = as(ConstPhysProp, (TVScale(Kshfg) ∘ TVExp(),))))
     return tr
 end
 """
@@ -72,6 +72,7 @@ This small function runs
 ```
 fitprm = transform(tr, fitlog)
 new_params = setproperties(po, fitprm)
+!isnothing(badprms) && badprms(new_params) && return NaN
 prob = ODEProblem(new_params; tspan=(0.0, 1000.0))
 sol = solve(prob, Rodas4(autodiff=AutoForwardDiff(chunksize=2)); saveat, kwargs...)
 ```

src/pikal_model.jl

@@ -144,12 +144,6 @@ function ParamObjPikal(tuptup)
     return ParamObjPikal(tuptup[1]..., tuptup[2]..., tuptup[3]...)
 end
 
-# This constructor catches if Rp is passed as a tuple or NamedTuple and constructs an appropriate object
-function ParamObjPikal(Rp::Union{NamedTuple, Tuple},
-    hf0, csolid, ρsolution, Kshf, Av, Ap, pch, Tsh)
-    return ParamObjPikal(RpFormFit(Rp...), hf0, csolid, ρsolution, Kshf, Av, Ap, pch, Tsh)
-end
-
 function Base.getindex(p::ParamObjPikal, i::Int)
     if i == 1
         return (p.Rp, p.hf0, p.csolid, p.ρsolution)

src/rf_lumcap_model.jl

@@ -182,11 +182,6 @@ function ParamObjRF(tuptup::Tuple)
 end
 Base.size(po::ParamObjRF) = (6,)
 
-# This constructor catches if Rp is passed as a tuple or NamedTuple and constructs an appropriate object
-function ParamObjRF(Rp::Union{NamedTuple, Tuple}, args...)
-    return ParamObjRF(RpFormFit(Rp...), args...)
-end
-
 function Base.getindex(po::ParamObjRF, i)
     if i == 1
         return (po.Rp, po.hf0, po.csolid, po.ρsolution)

src/structs.jl

@@ -4,6 +4,7 @@
     A1
     A2
 end
+RpFormFit(;R0, A1, A2) = RpFormFit(R0, A1, A2)
 @doc """
 A convenience type for dealing with the common functional form given to Rp and Kv.
 
@@ -267,13 +268,3 @@ function Base.:(==)(p1::PrimaryDryFit, p2::PrimaryDryFit)
     cond6 = ismissing(p1.t_end) ? ismissing(p2.t_end) : (p1.t_end == p2.t_end)
     return all([cond1, cond2, cond3, cond4, cond5, cond6])
 end
-
-# Add a little bit of sugar to our transforms
-struct ConstWrapTV <: TransformVariables.ScalarTransform end
-TransformVariables.transform(::ConstWrapTV, x) = ConstPhysProp(x)
-TransformVariables.inverse(::ConstWrapTV, x) = x.value
-
-# These create method ambiguities with other TransformVariables methods
-# I don't think they were needed, but if so can be brought back with more specificity
-# TransformVariables.transform(t::TVScale{ConstPhysProp}, x) = ConstPhysProp(t.scale.value*x)
-# TransformVariables.inverse(t::TVScale{ConstPhysProp}, x) = x.value/t.scale.value

test/test_KRp_opt.jl

@@ -48,10 +48,11 @@ pdfit = PrimaryDryFit(t, T; t_end)
     obj = OptimizationFunction(obj_pd, AutoForwardDiff(chunksize=4))
     opt = solve(OptimizationProblem(obj, pg, pass), optalg)
     vals = transform(tr, opt.u)
+    @test all(opt.u .!= 0)
     @test vals.Kshf(pch(0)) ≈ Kshf(pch(0)) rtol=0.3
     @test vals.Rp.R0 ≈ R0 rtol=0.1
-    @test vals.Rp.A1 ≈ A1 rtol=0.1
-    @test vals.Rp.A2 ≈ A2 rtol=0.1
+    @test vals.Rp.A1 ≈ A1 rtol=0.3
+    @test vals.Rp.A2 ≈ A2 rtol=0.5
 end
 
 @testset "Only Rp" begin
@@ -66,7 +67,7 @@ end
     opt = solve(OptimizationProblem(obj, pg, pass), optalg)
     vals = transform(tr, opt.u)
     @test vals.Rp.R0 ≈ R0 rtol=0.1
-    @test vals.Rp.A1 ≈ A1 rtol=0.1
+    @test vals.Rp.A1 ≈ A1 rtol=0.2
     @test vals.Rp.A2 ≈ A2 rtol=0.5
 end
 

test/test_jet.jl

@@ -1,4 +1,4 @@
 using JET
 # Can't use `mode = :basic` in CI because the Plots.jl recipe system triggers lots of false 
 # positives for JET. Not a bad idea to periodically check that manually, though.
-test_package(LyoPronto; mode = :typo, target_defined_modules=true)
+test_package(LyoPronto; mode = :typo, target_modules=(LyoPronto,))