Update cost function and time formatting

Author: leburgel

src/OptimKit.jl

@@ -104,6 +104,12 @@ Also see [`GradientDescent`](@ref), [`ConjugateGradient`](@ref), [`LBFGS`](@ref)
 """
 function optimize end
 
+function format_time(t::Float64)
+    return t < 60 ? @sprintf("%.2f s", t) :
+           t < 2600 ? @sprintf("%.2f m", t / 60) :
+           @sprintf("%.2f h", t / 3600)
+end
+
 include("linesearches.jl")
 include("gd.jl")
 include("cg.jl")

src/cg.jl

@@ -98,7 +98,7 @@ function optimize(fg, x, alg::ConjugateGradient;
 
     numiter = 0
     verbosity >= 2 &&
-        @info @sprintf("CG: initializing with f = %.12f, ‖∇f‖ = %.4e", f, normgrad)
+        @info @sprintf("CG: initializing with f = %.12e, ‖∇f‖ = %.4e", f, normgrad)
     local xprev, gprev, Pgprev, ηprev
     while !(_hasconverged || _shouldstop)
         told = t
@@ -150,8 +150,8 @@ function optimize(fg, x, alg::ConjugateGradient;
             break
         end
         verbosity >= 3 &&
-            @info @sprintf("CG: iter %4d, Δt %7.2f s: f = %.12f, ‖∇f‖ = %.4e, α = %.2e, β = %.2e, nfg = %d",
-                           numiter, Δt, f, normgrad, α, β, nfg)
+            @info @sprintf("CG: iter %4d, Δt %s: f = %.12e, ‖∇f‖ = %.4e, α = %.2e, β = %.2e, nfg = %d",
+                           numiter, format_time(Δt), f, normgrad, α, β, nfg)
 
         # transport gprev, ηprev and vectors in Hessian approximation to x
         gprev = transport!(gprev, xprev, ηprev, α, x)
@@ -167,12 +167,12 @@ function optimize(fg, x, alg::ConjugateGradient;
     end
     if _hasconverged
         verbosity >= 2 &&
-            @info @sprintf("CG: converged after %d iterations and time %.2f s: f = %.12f, ‖∇f‖ = %.4e",
-                           numiter, t, f, normgrad)
+            @info @sprintf("CG: converged after %d iterations and time %s: f = %.12e, ‖∇f‖ = %.4e",
+                           numiter, format_time(t), f, normgrad)
     else
         verbosity >= 1 &&
-            @warn @sprintf("CG: not converged to requested tol after %d iterations and time %.2f s: f = %.12f, ‖∇f‖ = %.4e",
-                           numiter, t, f, normgrad)
+            @warn @sprintf("CG: not converged to requested tol after %d iterations and time %s: f = %.12e, ‖∇f‖ = %.4e",
+                           numiter, format_time(t), f, normgrad)
     end
     history = [fhistory normgradhistory]
     return x, f, g, numfg, history

src/gd.jl

@@ -76,7 +76,7 @@ function optimize(fg, x, alg::GradientDescent;
 
     numiter = 0
     verbosity >= 2 &&
-        @info @sprintf("GD: initializing with f = %.12f, ‖∇f‖ = %.4e", f, normgrad)
+        @info @sprintf("GD: initializing with f = %.12e, ‖∇f‖ = %.4e", f, normgrad)
     while !(_hasconverged || _shouldstop)
         told = t
         # compute new search direction
@@ -107,20 +107,20 @@ function optimize(fg, x, alg::GradientDescent;
             break
         end
         verbosity >= 3 &&
-            @info @sprintf("GD: iter %4d, Δt %7.2f s: f = %.12f, ‖∇f‖ = %.4e, α = %.2e, nfg = %d",
-                           numiter, Δt, f, normgrad, α, nfg)
+            @info @sprintf("GD: iter %4d, Δt %s: f = %.12e, ‖∇f‖ = %.4e, α = %.2e, nfg = %d",
+                           numiter, format_time(Δt), f, normgrad, α, nfg)
 
         # increase α for next step
         α = 2 * α
     end
     if _hasconverged
         verbosity >= 2 &&
-            @info @sprintf("GD: converged after %d iterations and time %.2f s: f = %.12f, ‖∇f‖ = %.4e",
-                           numiter, t, f, normgrad)
+            @info @sprintf("GD: converged after %d iterations and time %s: f = %.12e, ‖∇f‖ = %.4e",
+                           numiter, format_time(t), f, normgrad)
     else
         verbosity >= 1 &&
-            @warn @sprintf("GD: not converged to requested tol after %d iterations and time %.2f s: f = %.12f, ‖∇f‖ = %.4e",
-                           numiter, t, f, normgrad)
+            @warn @sprintf("GD: not converged to requested tol after %d iterations and time %s: f = %.12e, ‖∇f‖ = %.4e",
+                           numiter, format_time(t), f, normgrad)
     end
     history = [fhistory normgradhistory]
     return x, f, g, numfg, history

src/lbfgs.jl

@@ -80,7 +80,7 @@ function optimize(fg, x, alg::LBFGS;
     H = LBFGSInverseHessian(m, TangentType[], TangentType[], ScalarType[])
 
     verbosity >= 2 &&
-        @info @sprintf("LBFGS: initializing with f = %.12f, ‖∇f‖ = %.4e", f, normgrad)
+        @info @sprintf("LBFGS: initializing with f = %.12e, ‖∇f‖ = %.4e", f, normgrad)
 
     while !(_hasconverged || _shouldstop)
         told = t
@@ -127,8 +127,8 @@ function optimize(fg, x, alg::LBFGS;
             break
         end
         verbosity >= 3 &&
-            @info @sprintf("LBFGS: iter %4d, Δt %7.2f: f = %.12f, ‖∇f‖ = %.4e, α = %.2e, m = %d, nfg = %d",
-                           numiter, Δt, f, normgrad, α, length(H), nfg)
+            @info @sprintf("LBFGS: iter %4d, Δt %s: f = %.12e, ‖∇f‖ = %.4e, α = %.2e, m = %d, nfg = %d",
+                           numiter, format_time(Δt), f, normgrad, α, length(H), nfg)
 
         # transport gprev, ηprev and vectors in Hessian approximation to x
         gprev = transport!(gprev, xprev, ηprev, α, x)
@@ -192,12 +192,12 @@ function optimize(fg, x, alg::LBFGS;
     end
     if _hasconverged
         verbosity >= 2 &&
-            @info @sprintf("LBFGS: converged after %d iterations and time %.2f s: f = %.12f, ‖∇f‖ = %.4e",
-                           numiter, t, f, normgrad)
+            @info @sprintf("LBFGS: converged after %d iterations and time %s: f = %.12e, ‖∇f‖ = %.4e",
+                           numiter, format_time(t), f, normgrad)
     else
         verbosity >= 1 &&
-            @warn @sprintf("LBFGS: not converged to requested tol after %d iterations and time %.2f s: f = %.12f, ‖∇f‖ = %.4e",
-                           numiter, t, f, normgrad)
+            @warn @sprintf("LBFGS: not converged to requested tol after %d iterations and time %s: f = %.12e, ‖∇f‖ = %.4e",
+                           numiter, format_time(t), f, normgrad)
     end
     history = [fhistory normgradhistory]
     return x, f, g, numfg, history