Merge pull request #84 from ACEsuit/co/valsvd

Add svd with validation set

Author: cortner

src/asp.jl

@@ -66,20 +66,23 @@ end
 function solve(solver::ASP, A, y, Aval=A, yval=y)
     # Apply preconditioning
     AP = A / solver.P
+    AvalP = Aval / solver.P
 
     tracer = asp_homotopy(AP, y; solver.params...)
 
     q = length(tracer) 
-    every = max(1, q ÷ solver.nstore)
-    istore = unique([1:every:q; q])
-    new_tracer = [ (solution = solver.P \ tracer[i][1], λ = tracer[i][2], σ = 0.0 ) 
+    every = max(1, q / solver.nstore)
+    istore = unique(round.(Int, [1:every:q; q]))
+    new_tracer = [ (solution = tracer[i][1], λ = tracer[i][2], σ = 0.0 ) 
                    for i in istore ]
 
     if solver.tsvd  # Post-processing if tsvd is true
-        post = post_asp_tsvd(new_tracer, A, y, Aval, yval)
-        new_post = [ (solution = p.θ, λ = p.λ, σ = p.σ) for p in post ]
+        post = post_asp_tsvd(new_tracer, AP, y, AvalP, yval)
+        new_post = [ (solution = solver.P \ p.θ, λ = p.λ, σ = p.σ) 
+                     for p in post ]
     else
-        new_post = new_tracer 
+        new_post = [ (solution = solver.P \ p.solution, λ = p.λ, σ = 0.0) 
+                     for p in new_tracer ]
     end
 
     xs, in = select_solution(new_post, solver, Aval, yval)
@@ -124,34 +127,6 @@ function select_solution(tracer, solver, A, y)
 end
 
 
-
-using SparseArrays
-
-function solve_tsvd(At, yt, Av, yv) 
-   Ut, Σt, Vt = svd(At); zt = Ut' * yt
-   Qv, Rv = qr(Av); zv = Matrix(Qv)' * yv
-   @assert issorted(Σt, rev=true)
-
-   Rv_Vt = Rv * Vt
-
-   θv = zeros(size(Av, 2))
-   θv[1] = zt[1] / Σt[1] 
-   rv = Rv_Vt[:, 1] * θv[1] - zv 
-
-   tsvd_errs = Float64[] 
-   push!(tsvd_errs, norm(rv))
-
-   for k = 2:length(Σt)
-      θv[k] = zt[k] / Σt[k]
-      rv += Rv_Vt[:, k] * θv[k]
-      push!(tsvd_errs, norm(rv))
-   end
-
-   imin = argmin(tsvd_errs)
-   θv[imin+1:end] .= 0
-   return Vt * θv, Σt[imin]
-end
-
 function post_asp_tsvd(path, At, yt, Av, yv) 
    Qt, Rt = qr(At); zt = Matrix(Qt)' * yt
    Qv, Rv = qr(Av); zv = Matrix(Qv)' * yv
@@ -166,14 +141,4 @@ function post_asp_tsvd(path, At, yt, Av, yv)
    end
 
    return _post.(path)
-
-#    post = [] 
-#    for (θ, λ) in path
-#       if isempty(θ.nzind); push!(post, (θ = θ, λ = λ, σ = Inf)); continue; end  
-#       inz = θ.nzind 
-#       θ1, σ = solve_tsvd(Rt[:, inz], zt, Rv[:, inz], zv)
-#       θ2 = copy(θ); θ2[inz] .= θ1
-#       push!(post, (θ = θ2, λ = λ, σ = σ))
-#    end 
-#    return identity.(post)
 end

src/solvers.jl

@@ -196,3 +196,37 @@ function solve(solver::TruncatedSVD, A, y)
     return Dict{String, Any}("C" => solver.P \ θP)
 end
 
+
+# ------------ Truncated SVD with tol specified by validation set ------------ 
+
+function solve_tsvd(At, yt, Av, yv) 
+    Ut, Σt, Vt = svd(At); zt = Ut' * yt
+    Qv, Rv = qr(Av); zv = Matrix(Qv)' * yv
+    @assert issorted(Σt, rev=true)
+ 
+    Rv_Vt = Rv * Vt
+ 
+    θv = zeros(size(Av, 2))
+    θv[1] = zt[1] / Σt[1] 
+    rv = Rv_Vt[:, 1] * θv[1] - zv 
+ 
+    tsvd_errs = Float64[] 
+    push!(tsvd_errs, norm(rv))
+ 
+    for k = 2:length(Σt)
+       θv[k] = zt[k] / Σt[k]
+       rv += Rv_Vt[:, k] * θv[k]
+       push!(tsvd_errs, norm(rv))
+    end
+ 
+    imin = argmin(tsvd_errs)
+    θv[imin+1:end] .= 0
+    return Vt * θv, Σt[imin]
+ end
+
+
+function solve(solver::TruncatedSVD, At, yt, Av, yv)
+    # make a function barrier because solver.P is not inferred 
+    θ, σ = solve_tsvd(At / solver.P, yt, Av / solver.P, yv)
+    return Dict{String, Any}("C" => solver.P \ θ, "σ" => σ)
+end

test/test_asp.jl

@@ -1,6 +1,5 @@
 using ACEfit
 using LinearAlgebra, Random, Test 
-using Random
 
 ##
 
@@ -29,6 +28,12 @@ Av = A[val_indices,:]
 yt = y[train_indices]
 yv = y[val_indices]
 
+for (nstore, n1) in [ (20, 21), (100, 101), (200, 165)]
+   solver = ACEfit.ASP(P=I, select = :final, nstore = nstore, loglevel=0, traceFlag=true)
+   results = ACEfit.solve(solver, A, y)
+   @test length(results["path"]) == n1 
+end
+
 for (select, tolr, tolc) in [ (:final, 10*epsn, 1), 
                              ( (:byerror,1.3), 10*epsn, 1), 
                             ( (:bysize,73), 1, 10) ]

test/test_linearsolvers.jl

@@ -1,8 +1,5 @@
 
-using ACEfit
-using LinearAlgebra, Random, Test 
-using Random
-using PythonCall
+using ACEfit, LinearAlgebra, Random, Test, PythonCall
 
 ##
 
@@ -168,3 +165,19 @@ C = results["C"]
 @test norm(A * C - y) < 10 * epsn 
 @test norm(C - c_ref) < 1
 
+
+##
+
+@info("Truncated SVD with validation")
+solver = ACEfit.TruncatedSVD(; rtol =  0.0)
+At = A[1:8000, :] 
+yt = y[1:8000]
+Av = A[8001:end, :]
+yv = y[8001:end]
+results_v = ACEfit.solve(solver, At, yt, Av, yv)
+@show err_v = norm(Av * results_v["C"] - yv)
+@show err = norm(Av * results["C"] - yv)
+@test err_v <= err 
+@show norm(results_v["C"] - c_ref)
+@show norm(results["C"] - c_ref)
+@test norm(results_v["C"] - c_ref) < 1e-2