Require Julia 0.7 (#127)

* Move to Julia 0.7
* Explicitly call `convert(T, ::Number)` rather than `T(::Number)`

.travis.yml

@@ -4,7 +4,7 @@ os:
   - linux
   - osx
 julia:
-  - 0.6
+  - 0.7
   - nightly
 matrix:
   allow_failures:

REQUIRE

@@ -1,5 +1,4 @@
-julia 0.6
-NLSolversBase 5.0
+julia 0.7-beta2
+NLSolversBase 7.0
 Parameters
 NaNMath
-Compat

appveyor.yml

@@ -1,7 +1,7 @@
 environment:
   matrix:
-  - JULIA_URL: "https://julialang-s3.julialang.org/bin/winnt/x86/0.6/julia-0.6-latest-win32.exe"
-  - JULIA_URL: "https://julialang-s3.julialang.org/bin/winnt/x64/0.6/julia-0.6-latest-win64.exe"
+  - JULIA_URL: "https://julialang-s3.julialang.org/bin/winnt/x86/0.7/julia-0.7-latest-win32.exe"
+  - JULIA_URL: "https://julialang-s3.julialang.org/bin/winnt/x64/0.7/julia-0.7-latest-win64.exe"
 matrix:
   allow_failures:
     - JULIA_URL: "https://julialangnightlies-s3.julialang.org/bin/winnt/x86/julia-latest-win32.exe"

docs/generate.jl

@@ -3,12 +3,13 @@ import Literate
 
 # TODO: Remove items from `SKIPFILE` as soon as they run on the latest
 # stable `Optim` (or other dependency)
-#ONLYSTATIC = ["optim_linesearch.jl", "optim_initialstep.jl"]
-ONLYSTATIC = []
+ONLYSTATIC = ["optim_linesearch.jl", "optim_initialstep.jl"]
+#ONLYSTATIC = []
 
 EXAMPLEDIR = joinpath(@__DIR__, "src", "examples")
 GENERATEDDIR = joinpath(@__DIR__, "src", "examples", "generated")
-for example in filter!(r"\.jl$", readdir(EXAMPLEDIR))
+myfilter(str) = occursin(r"\.jl$", str)
+for example in filter!(myfilter, readdir(EXAMPLEDIR))
     input = abspath(joinpath(EXAMPLEDIR, example))
     script = Literate.script(input, GENERATEDDIR)
     code = strip(read(script, String))

docs/make.jl

@@ -11,7 +11,7 @@ makedocs(
     format = :html,
     sitename = "LineSearches.jl",
     doctest = false,
-    # strict = VERSION.minor == 6 && sizeof(Int) == 8, # only strict mode on 0.6 and Int64
+    # strict = VERSION.minor == 7 && sizeof(Int) == 8, # only strict mode on 0.7 and Int64
     strict = false,
     pages = Any[
         "Home" => "index.md",
@@ -27,7 +27,7 @@ makedocs(
 deploydocs(
     repo = "github.com/JuliaNLSolvers/LineSearches.jl.git",
     target = "build",
-    julia = "0.6", # deploy from release bot
+    julia = "0.7", # deploy from release bot
     deps = nothing,
     make = nothing,
 )

docs/src/examples/customoptimizer.jl

@@ -28,11 +28,11 @@ function gdoptimize(f, g!, fg!, x0::AbstractArray{T}, linesearch,
     ϕ(α) = f(x .+ α.*s)
     function dϕ(α)
         g!(gvec, x .+ α.*s)
-        return vecdot(gvec, s)
+        return dot(gvec, s)
     end
     function ϕdϕ(α)
         phi = fg!(gvec, x .+ α.*s)
-        dphi = vecdot(gvec, s)
+        dphi = dot(gvec, s)
         return (phi, dphi)
     end
 
@@ -100,10 +100,10 @@ ls = BackTracking(order=3)
 fx_bt3, x_bt3, iter_bt3 = gdoptimize(f, g!, fg!, x0, ls)
 
 ## Test the results                #src
-using Base.Test                    #src
+using Test                         #src
 @test fx_bt3 < 1e-12               #src
 @test iter_bt3 < 10000             #src
-@test x_bt3 ≈ [1.0, 1.0] atol=1e-7 #src
+@test x_bt3 ≈ [1.0, 1.0] atol=2e-7 #src
 
 # Interestingly, the `StrongWolfe` line search converges in one iteration, whilst
 # all the other algorithms take thousands of iterations.

docs/src/examples/optim_initialstep.jl

@@ -1,11 +1,5 @@
 # # Optim initial step length guess
 #
-#src TODO: Find a way to run these with Literate when deploying via Travis
-#src TODO: This file must currently be run locally and not on CI, and then
-#src TODO: the md file must be copied over to the correct directory.
-#src TODO: The reason is that there may be breaking changes between Optim and LineSearches,
-#src TODO: so we don't want that to mess up JuliaCIBot
-#-
 #-
 #md # !!! tip
 #md #     This example is also available as a Jupyter notebook:
@@ -34,7 +28,7 @@ res_hz = Optim.optimize(prob.f, prob.g!, prob.h!, prob.initial_x, method=algo_hz
 # From the result we see that this has reduced the number of function and gradient calls, but increased the number of iterations.
 
 ## Test the results                                       #src
-using Base.Test                                           #src
+using Test                                                #src
 @test Optim.f_calls(res_hz) < Optim.f_calls(res_st)       #src
 @test Optim.g_calls(res_hz) < Optim.g_calls(res_st)       #src
 @test Optim.iterations(res_hz) > Optim.iterations(res_st) #src

docs/src/examples/optim_linesearch.jl

@@ -1,10 +1,5 @@
 # # Optim line search
 #
-#src TODO: Find a way to run these with Literate when deploying via Travis
-#src TODO: This file must currently be run locally and not on CI, and then
-#src TODO: the md file must be copied over to the correct directory.
-#src TODO: The reason is that there may be breaking changes between Optim and LineSearches,
-#src TODO: so we don't want that to mess up JuliaCIBot
 #-
 #md # !!! tip
 #md #     This example is also available as a Jupyter notebook:
@@ -32,7 +27,7 @@ algo_bt3 = Newton(linesearch = BackTracking(order=3))
 res_bt3 = Optim.optimize(prob.f, prob.g!, prob.h!, prob.initial_x, method=algo_bt3)
 
 
-## Test the results                                #src
-using Base.Test                                    #src
+## Test the results                                  #src
+using Test                                           #src
 @test Optim.f_calls(res_bt3) < Optim.f_calls(res_hz) #src
 @test Optim.g_calls(res_bt3) < Optim.g_calls(res_hz) #src

src/LineSearches.jl

@@ -1,12 +1,9 @@
-isdefined(Base, :__precompile__) && __precompile__()
+__precompile__()
 
 module LineSearches
 
-using   Compat,
-        Compat.LinearAlgebra,
-        Compat.Distributed,
-        Compat.Printf
-
+using Printf
+import LinearAlgebra: dot, norm
 using Parameters, NaNMath
 
 import NLSolversBase
@@ -38,7 +35,7 @@ function make_ϕdϕ(df, x_new, x, s)
         NLSolversBase.value_gradient!(df, x_new)
 
         # Calculate ϕ(a_i), ϕ'(a_i)
-        NLSolversBase.value(df), real(vecdot(NLSolversBase.gradient(df), s))
+        NLSolversBase.value(df), real(dot(NLSolversBase.gradient(df), s))
     end
     ϕdϕ
 end
@@ -51,7 +48,7 @@ function make_ϕ_dϕ(df, x_new, x, s)
         NLSolversBase.gradient!(df, x_new)
 
         # Calculate ϕ'(a_i)
-        real(vecdot(NLSolversBase.gradient(df), s))
+        real(dot(NLSolversBase.gradient(df), s))
     end
     make_ϕ(df, x_new, x, s), dϕ
 end
@@ -64,7 +61,7 @@ function make_ϕ_dϕ_ϕdϕ(df, x_new, x, s)
         NLSolversBase.gradient!(df, x_new)
 
         # Calculate ϕ'(a_i)
-        real(vecdot(NLSolversBase.gradient(df), s))
+        real(dot(NLSolversBase.gradient(df), s))
     end
     function ϕdϕ(α)
         # Move a distance of alpha in the direction of s
@@ -74,7 +71,7 @@ function make_ϕ_dϕ_ϕdϕ(df, x_new, x, s)
         NLSolversBase.value_gradient!(df, x_new)
 
         # Calculate ϕ'(a_i)
-        NLSolversBase.value(df), real(vecdot(NLSolversBase.gradient(df), s))
+        NLSolversBase.value(df), real(dot(NLSolversBase.gradient(df), s))
     end
     make_ϕ(df, x_new, x, s), dϕ, ϕdϕ
 end
@@ -87,7 +84,7 @@ function make_ϕ_ϕdϕ(df, x_new, x, s)
         NLSolversBase.value_gradient!(df, x_new)
 
         # Calculate ϕ'(a_i)
-        NLSolversBase.value(df), real(vecdot(NLSolversBase.gradient(df), s))
+        NLSolversBase.value(df), real(dot(NLSolversBase.gradient(df), s))
     end
     make_ϕ(df, x_new, x, s), ϕdϕ
 end

src/backtracking.jl

@@ -29,7 +29,7 @@ function (ls::BackTracking)(df::AbstractObjective, x::AbstractArray{T}, s::Abstr
         dϕ_0 = dϕ(Tα(0))
     end
 
-    α_0 = min(α_0, min(alphamax, ls.maxstep / vecnorm(s, Inf)))
+    α_0 = min(α_0, min(alphamax, ls.maxstep / norm(s, Inf)))
     ls(ϕ, α_0, ϕ_0, dϕ_0)
 end
 

src/hagerzhang.jl

@@ -111,18 +111,19 @@ function (ls::HagerZhang)(ϕ, ϕdϕ,
     @unpack delta, sigma, alphamax, rho, epsilon, gamma,
             linesearchmax, psi3, display, mayterminate = ls
 
+    zeroT = convert(T, 0)
 
     if !(isfinite(phi_0) && isfinite(dphi_0))
         throw(ArgumentError("Value and slope at step length = 0 must be finite."))
     end
-    if dphi_0 >= T(0)
+    if dphi_0 >= zeroT
         throw(ArgumentError("Search direction is not a direction of descent."))
     end
 
     # Prevent values of x_new = x+αs that are likely to make
     # ϕ(x_new) infinite
     iterfinitemax::Int = ceil(Int, -log2(eps(T)))
-    alphas = [T(0)] # for bisection
+    alphas = [zeroT] # for bisection
     values = [phi_0]
     slopes = [dphi_0]
     if display & LINESEARCH > 0
@@ -131,7 +132,7 @@ function (ls::HagerZhang)(ϕ, ϕdϕ,
 
 
     phi_lim = phi_0 + epsilon * abs(phi_0)
-    @assert c > T(0)
+    @assert c > zeroT
     @assert isfinite(c) && c <= alphamax
     phi_c, dphi_c = ϕdϕ(c)
     iterfinite = 1
@@ -142,9 +143,9 @@ function (ls::HagerZhang)(ϕ, ϕdϕ,
         phi_c, dphi_c = ϕdϕ(c)
     end
     if !(isfinite(phi_c) && isfinite(dphi_c))
-        warn("Failed to achieve finite new evaluation point, using alpha=0")
+        @warn("Failed to achieve finite new evaluation point, using alpha=0")
         mayterminate[] = false # reset in case another initial guess is used next
-        return T(0.0), ϕ(T(0.0)) # phi_0
+        return zeroT, ϕ(zeroT) # phi_0
     end
     push!(alphas, c)
     push!(values, phi_c)
@@ -175,7 +176,7 @@ function (ls::HagerZhang)(ϕ, ϕdϕ,
                     ", phi_c = ", phi_c,
                     ", dphi_c = ", dphi_c)
         end
-        if dphi_c >= T(0)
+        if dphi_c >= zeroT
             # We've reached the upward slope, so we have b; examine
             # previous values to find a
             ib = length(alphas)
@@ -191,7 +192,7 @@ function (ls::HagerZhang)(ϕ, ϕdϕ,
             # have crested over the peak. Use bisection.
             ib = length(alphas)
             ia = ib - 1
-            if c ≉  alphas[ib] || slopes[ib] >= T(0)
+            if c ≉  alphas[ib] || slopes[ib] >= zeroT
                 error("c = ", c)
             end
             # ia, ib = bisect(phi, lsr, ia, ib, phi_lim) # TODO: Pass options
@@ -226,7 +227,7 @@ function (ls::HagerZhang)(ϕ, ϕdϕ,
             if !(isfinite(phi_c) && isfinite(dphi_c))
                 mayterminate[] = false # reset in case another initial guess is used next
                 return cold, ϕ(cold)
-            elseif dphi_c < T(0) && c == alphamax
+            elseif dphi_c < zeroT && c == alphamax
                 # We're on the edge of the allowed region, and the
                 # value is still decreasing. This can be due to
                 # roundoff error in barrier penalties, a barrier
@@ -352,7 +353,8 @@ function secant2!(ϕdϕ,
     dphi_a = slopes[ia]
     dphi_b = slopes[ib]
     T = eltype(slopes)
-    if !(dphi_a < T(0) && dphi_b >= T(0))
+    zeroT = convert(T, 0)
+    if !(dphi_a < zeroT && dphi_b >= zeroT)
         error(string("Search direction is not a direction of descent; ",
                      "this error may indicate that user-provided derivatives are inaccurate. ",
                       @sprintf "(dphi_a = %f; dphi_b = %f)" dphi_a dphi_b))
@@ -436,10 +438,11 @@ function update!(ϕdϕ,
     a = alphas[ia]
     b = alphas[ib]
     T = eltype(slopes)
+    zeroT = convert(T, 0)
     # Debugging (HZ, eq. 4.4):
-    @assert slopes[ia] < T(0)
+    @assert slopes[ia] < zeroT
     @assert values[ia] <= phi_lim
-    @assert slopes[ib] >= T(0)
+    @assert slopes[ib] >= zeroT
     @assert b > a
     c = alphas[ic]
     phi_c = values[ic]
@@ -456,7 +459,7 @@ function update!(ϕdϕ,
     if c < a || c > b
         return ia, ib #, 0, 0  # it's out of the bracketing interval
     end
-    if dphi_c >= T(0)
+    if dphi_c >= zeroT
         return ia, ic #, 0, 0  # replace b with a closer point
     end
     # We know dphi_c < 0. However, phi may not be monotonic between a
@@ -485,9 +488,10 @@ function bisect!(ϕdϕ,
     a = alphas[ia]
     b = alphas[ib]
     # Debugging (HZ, conditions shown following U3)
-    @assert slopes[ia] < T(0)
+    zeroT = convert(T, 0)
+    @assert slopes[ia] < zeroT
     @assert values[ia] <= phi_lim
-    @assert slopes[ib] < T(0)       # otherwise we wouldn't be here
+    @assert slopes[ib] < zeroT       # otherwise we wouldn't be here
     @assert values[ib] > phi_lim
     @assert b > a
     while b - a > eps(b)
@@ -503,7 +507,7 @@ function bisect!(ϕdϕ,
         push!(slopes, gphi)
 
         id = length(alphas)
-        if gphi >= T(0)
+        if gphi >= zeroT
             return ia, id # replace b, return
         end
         if phi_d <= phi_lim