update

.gitignore

@@ -1,5 +1,5 @@
 src/temp.jl
-#test/precompile/
+test/precompile/
 test/devtest.jl
 examples/
 docs/build/

Project.toml

@@ -1,7 +1,7 @@
 name = "Metida"
 uuid = "a1dec852-9fe5-11e9-361f-8d9fde67cfa2"
 authors = ["Vladimir Arnautov <[email protected]>"]
-version = "0.1.9"
+version = "0.1.10"
 
 [deps]
 CategoricalArrays = "324d7699-5711-5eae-9e2f-1d82baa6b597"
@@ -23,7 +23,7 @@ Optim = "0.19, 0.20, 0.21, 0.22, 1.0, 1.1, 1.2"
 StatsBase = "0.29, 0.30, 0.31, 0.32, 0.33"
 StatsModels = "0.6"
 Tables = "1"
-julia = "1.1, 1.2, 1.3, 1.4, 1.5"
+julia = "1.1, 1.2, 1.3, 1.4, 1.5, 1.6"
 
 [extras]
 CSV = "336ed68f-0bac-5ca0-87d4-7b16caf5d00b"

README.md

@@ -29,7 +29,7 @@ categorical!(df, :formulation);
 
 lmm = LMM(@formula(var~sequence+period+formulation), df;
 random = VarEffect(@covstr(formulation), CSH),
-repeated = VarEffect(@covstr(formulation), VC),
+repeated = VarEffect(@covstr(formulation), DIAG),
 subject = :subject)
 
 fit!(lmm)

src/fit.jl

@@ -5,7 +5,7 @@
 
 Fit LMM model.
 """
-function fit!(lmm::LMM{T}; verbose::Symbol = :auto, varlinkf = :exp, rholinkf = :sigm) where T
+function fit!(lmm::LMM{T}; verbose::Symbol = :auto, varlinkf = :exp, rholinkf = :sigm, aifirst = false) where T
 
     #Make varlink function
     fv  = varlinkvec(lmm.covstr.ct)
@@ -41,32 +41,35 @@ function fit!(lmm::LMM{T}; verbose::Symbol = :auto, varlinkf = :exp, rholinkf =
     for i = 1:length(θ)
         if lmm.covstr.ct[i] == :rho θ[i] = 0.0 end
     end
-
+    push!(lmm.log, "Initial θ: "*string(θ))
     #Initial step with modified Newton method
     ############################################################################
-    aif(x) = optfunc(lmm, x)[4]
-    grf(x) = optfunc(lmm, x)[1]
-    ai = ForwardDiff.hessian(aif, θ)
-    gr = ForwardDiff.gradient(grf, θ)
-    println("vec: ", θ)
-    try
-        θ .-= (inv(ai) ./4 )*gr
-    catch
-        θ .-= (pinv(ai) ./4 )*gr
-    end
-    for i = 1:length(θ)
-        if lmm.covstr.ct[i] == :rho
-            if θ[i] > 0.99
-                θ[i] = 0.9
-            elseif θ[i] < 0.0
-                θ[i] = 0.0
+    if aifirst
+        aif(x) = optfunc(lmm, x)[4]
+        grf(x) = optfunc(lmm, x)[1]
+        ai = ForwardDiff.hessian(aif, θ)
+        gr = ForwardDiff.gradient(grf, θ)
+    #println("vec: ", θ)
+        try
+            θ .-= (inv(ai) ./4 )*gr
+        catch
+            θ .-= (pinv(ai) ./4 )*gr
+        end
+        for i = 1:length(θ)
+            if lmm.covstr.ct[i] == :rho
+                if θ[i] > 0.99
+                    θ[i] = 0.9
+                elseif θ[i] < 0.0
+                    θ[i] = 0.0
+                end
+            else
+                if θ[i] < 0.01 θ[i] = initθ / 2 end
+                if θ[i] > initθ*1.25 θ[i] = initθ*1.25 end
             end
-        else
-            if θ[i] < 0.01 θ[i] = initθ / 2 end
-            if θ[i] > initθ*1.25 θ[i] = initθ*1.25 end
         end
+        push!(lmm.log, "First step with AI like method, θ: "*string(θ))
     end
-    println("new: ", θ)
+    #println("new: ", θ)
     #varlinkvecapply!(θ, fvr)
     varlinkrvecapply2!(θ, lmm.covstr.ct)
     #Twice differentiable object

src/reml.jl

@@ -38,7 +38,7 @@ end
 """
     -2 log Restricted Maximum Likelihood; β calculation inside
 """
-function reml_sweep_β(lmm::LMM{T2}, θ::Vector{T}) where T <: Number where T2 <: Number
+function reml_sweep_β(lmm::LMM{T2}, @nospecialize θ::Vector{T}) where T <: Number where T2 <: Number
     n::Int        = length(lmm.data.block)
     N::Int        = length(lmm.data.yv)
     G::Matrix{T}  = gmat_base(θ, lmm.covstr)
@@ -95,7 +95,7 @@ function reml_sweep_β(lmm::LMM{T2}, θ::Vector{T}) where T <: Number where T2 <
     return   θ₁ + logdetθ₂ + θ₃ + c, β, θ₂, θ₃
 end
 
-function reml_sweep_β2(lmm::LMM{T2}, θ::Vector{T}) where T <: Number where T2 <: Number
+function reml_sweep_β2(lmm::LMM{T2}, @nospecialize θ::Vector{T}) where T <: Number where T2 <: Number
     n::Int        = length(lmm.data.block)
     N::Int        = length(lmm.data.yv)
     #G::Matrix{T}  = gmat_base(θ, lmm.covstr)
@@ -150,7 +150,7 @@ function reml_sweep_β2(lmm::LMM{T2}, θ::Vector{T}) where T <: Number where T2
     return   θ₁ + logdetθ₂ + θ₃ + c, β, θ₂, θ₃
 end
 
-function reml_sweep_β3(lmm::LMM{T2}, θ::Vector{T}) where T <: Number where T2 <: Number
+function reml_sweep_β3(lmm::LMM{T2}, @nospecialize θ::Vector{T}) where T <: Number where T2 <: Number
     n::Int        = length(lmm.data.block)
     N::Int        = length(lmm.data.yv)
     G::Matrix{T}  = gmat_base(θ, lmm.covstr)

test/berds.jl

@@ -14,7 +14,7 @@ remlsb = [536.2011487949176,-28.588171019268458,436.5116765557768,316.0505671599
 SPSS
 REML G - CSH, R - DIAG
 13 - 2087.4810139727 / 2087.481017533285
-15 - / 2087.481017533285
+15 - / 2087.481017533285 / 2087.4810192819373
 16 - 323.99766668073 / 323.9976737508006
 29 - / 26.96606070210349
 30 - 14.944038069158
@@ -37,19 +37,25 @@ for i = 1:30
     df.lnpk = log.(df.PK)
 
     @testset "  Test $(i)" begin
+        atol=1E-6
+
         lmm = Metida.LMM(@formula(lnpk~sequence+period+treatment), df;
-        random = Metida.VarEffect(Metida.@covstr(treatment), Metida.CSH),
-        repeated = Metida.VarEffect(Metida.@covstr(treatment), Metida.DIAG),
+        random = Metida.VarEffect(Metida.@covstr(1), Metida.SI),
         subject = :subject
         )
         Metida.fit!(lmm)
-        @test lmm.result.reml ≈ remls[i] atol=1E-6
+        @test lmm.result.reml ≈ remlsb[i] atol=atol
 
+
+        if i == 13 || i == 15 atol = 1E-4 end
         lmm = Metida.LMM(@formula(lnpk~sequence+period+treatment), df;
-        random = Metida.VarEffect(Metida.@covstr(1), Metida.SI),
+        random = Metida.VarEffect(Metida.@covstr(treatment), Metida.CSH),
+        repeated = Metida.VarEffect(Metida.@covstr(treatment), Metida.DIAG),
         subject = :subject
         )
         Metida.fit!(lmm)
-        @test lmm.result.reml ≈ remlsb[i] atol=1E-6
+        @test lmm.result.reml ≈ remls[i] atol=atol
+
+
     end
 end

test/snoopy.jl

@@ -15,3 +15,13 @@ end
 data = SnoopCompile.read("$path/precompile/metida_compiles.log")
 pc = SnoopCompile.parcel(reverse!(data[2]))
 SnoopCompile.write("$path/precompile", pc)
+
+
+using SnoopCompileCore
+invalidations = @snoopr begin
+    using Metida
+end
+using SnoopCompile
+ui = uinvalidated(invalidations)
+trees = invalidation_trees(invalidations)
+ftrees = filtermod(Metida, trees)

test/test.jl

@@ -160,3 +160,4 @@ end
 include("ar.jl")
 include("lme4.jl")
 include("norand.jl")
+include("berds.jl")