wip

experiments/dual_svm/runme.jl

@@ -11,8 +11,8 @@ using AdaProx
 using Random
 using Plots
 using LaTeXStrings
-using ProximalCore
-using ProximalOperators: IndBox, IndZero
+using ProximalCore: IndZero
+using ProximalOperators: IndBox
 
 pgfplotsx()
 
@@ -21,15 +21,10 @@ struct Quadratic{TQ,Tq}
     q::Tq
 end
 
-function (f::Quadratic)(x)
+function AdaProx.eval_with_pullback(f::Quadratic, x)
     temp = f.Q * x
-    return 0.5 * dot(x, temp) + dot(x, f.q)
-end
-
-function ProximalCore.gradient!(grad, f::Quadratic, x)
-    temp = f.Q * x
-    grad .= temp + f.q
-    return 0.5 * dot(x, temp) + dot(x, f.q)
+    quadratic_pullback() = temp + f.q
+    return 0.5 * dot(x, temp) + dot(x, f.q), quadratic_pullback
 end
 
 function run_dsvm(
@@ -56,7 +51,7 @@ function run_dsvm(
     f = Quadratic(Q, q)
     g = IndBox(0.0, C)
     h = IndZero()
-    A = y'
+    A = Matrix(y')
 
     Lf = norm(Q)
     x0 = zeros(N)
@@ -117,7 +112,7 @@ function plot_residual(path)
     names_to_plot = []
     for name in ["Condat-Vu", "Malitsky-Pock", "AdaPDM"]
         matching_names = [k for k in keys(gb) if startswith(k.method, name)]
-        push!(names_to_plot, find_best(gb, matching_names, :norm_res, 1e-5, :grad_f_evals))
+        push!(names_to_plot, find_best(gb, matching_names, :norm_res, 1e-5, :f_evals))
     end
 
     fig = plot(
@@ -131,7 +126,7 @@ function plot_residual(path)
             continue
         end
         plot!(
-            gb[k][!, :grad_f_evals],
+            gb[k][!, :f_evals],
             gb[k][!, :norm_res],
             yaxis = :log,
             label = k.method,
@@ -143,7 +138,7 @@ end
 
 
 function main(;maxit = 10_000)
-    keys_to_log = [:method, :it, :grad_f_evals, :norm_res]
+    keys_to_log = [:method, :it, :f_evals, :norm_res]
 
     for C  in [0.1, 1]
         path = joinpath(@__DIR__, "svmguide3_C_$(C).jsonl")

experiments/least_absolute_deviation/runme.jl

@@ -15,6 +15,11 @@ using AdaProx
 
 pgfplotsx()
 
+function AdaProx.eval_with_pullback(f::Zero, x)
+    zero_pullback() = zero(x)
+    return f(x), zero_pullback
+end
+
 function run_least_absolute_deviation(
     filename,
     ::Type{T} = Float64;

experiments/nesterov_worst_case/runme.jl

@@ -16,28 +16,29 @@ struct WorstQuadratic{S,T}
     L::T
 end
 
-function (f::WorstQuadratic)(x)
+function AdaProx.eval_with_pullback(f::WorstQuadratic, x)
     s = x[1]^2 + x[f.k]^2
+
     for i in 1:(f.k-1)
         s += (x[i] - x[i+1])^2
     end
-    return (f.L / 4) * (s / 2 - x[1])
-end
 
-function ProximalCore.gradient!(grad, f::WorstQuadratic, x)
-    grad[1] = (f.L / 4) * (2 * x[1] - x[2] - 1)
-    for i in 2:(f.k-1)
-        grad[i] = (f.L / 4) * (2 * x[i] - x[i-1] - x[i+1])
+    function worst_quadratic_pullback()
+        grad = zero(x)
+        grad[1] = (f.L / 4) * (2 * x[1] - x[2] - 1)
+        for i in 2:(f.k-1)
+            grad[i] = (f.L / 4) * (2 * x[i] - x[i-1] - x[i+1])
+        end
+        grad[f.k] = (f.L / 4) * (2 * x[f.k] - x[f.k-1])
+        grad[(f.k+1):end] .= 0
+        return grad
     end
-    grad[f.k] = (f.L / 4) * (2 * x[f.k] - x[f.k-1])
-    grad[(f.k+1):end] .= 0
-    # since f is quadratic, f(x) = 1/2 <x, Q x> + <q, x>
-    # meaning \nabla f(x) = Q x + q
-    # and f(x) = (1/2) <\nabla f(x), x> + (1/2) <q, x>
-    # since q = -(L/4) e_1, we obtain the following
-    return dot(grad, x) / 2 - (f.L / 8) * x[1]
+
+    return (f.L / 4) * (s / 2 - x[1]), worst_quadratic_pullback
 end
 
+(f::WorstQuadratic)(x) = AdaProx.eval_with_pullback(f, x)[1]
+
 function run_nesterov_worst_case()
     k = 100
     n = 100

experiments/square_root_lasso/runme.jl

@@ -15,6 +15,11 @@ using AdaProx
 
 pgfplotsx()
 
+function AdaProx.eval_with_pullback(f::Zero, x)
+    zero_pullback() = zero(x)
+    return f(x), zero_pullback
+end
+
 function run_square_root_lasso(
     filename,
     ::Type{T} = Float64;