Merge pull request #37 from LAMPSPUC/dev

Version 0.1.4

Author: Giovanni3A

Project.toml

@@ -1,7 +1,7 @@
 name = "ApplicationDrivenLearning"
 uuid = "0856f1c8-ef17-4e14-9230-2773e47a789e"
 authors = ["Giovanni Amorim", "Joaquim Garcia"]
-version = "0.1.3"
+version = "0.1.4"
 
 [deps]
 BilevelJuMP = "485130c0-026e-11ea-0f1a-6992cd14145c"

src/optimizers/gradient.jl

@@ -34,6 +34,7 @@ function train_with_gradient!(
     verbose = get(params, :verbose, true)
     compute_cost_every = get(params, :compute_cost_every, 1)
     time_limit = get(params, :time_limit, Inf)
+    g_tol = get(params, :g_tol, 0)
 
     # init parameters
     start_time = time()
@@ -87,6 +88,17 @@ function train_with_gradient!(
 
         # check time limit reach
         if time() - start_time > time_limit
+            if verbose
+                println("Time limit reached.")
+            end
+            break
+        end
+
+        # check gradient tolerance
+        if maximum(abs.(dC)) < g_tol
+            if verbose
+                println("Gradient tolerance reached.")
+            end
             break
         end
 

src/optimizers/gradient_mpi.jl

@@ -17,14 +17,15 @@ function train_with_gradient_mpi!(
     compute_cost_every = get(params, :compute_cost_every, 1)
     mpi_finalize = get(params, :mpi_finalize, true)
     time_limit = get(params, :time_limit, Inf)
+    g_tol = get(params, :g_tol, 0)
 
     JQM.mpi_init()
 
     # init parameters
     start_time = time()
     is_done = false
     best_C = Inf
-    best_θ = []
+    best_θ = extract_params(model.forecast)
     curr_C = 0.0
     trace = Array{Float64}(undef, epochs)
     dCdz = Vector{Float32}(undef, size(model.policy_vars, 1))
@@ -121,6 +122,17 @@ function train_with_gradient_mpi!(
 
             # check time limit reach
             if time() - start_time > time_limit
+                if verbose
+                    println("Time limit reached.")
+                end
+                break
+            end
+
+            # check gradient tolerance
+            if maximum(abs.(dCdy)) < g_tol
+                if verbose
+                    println("Gradient tolerance reached.")
+                end
                 break
             end
 

src/predictive_model.jl

@@ -63,8 +63,9 @@ When only one network is passed as a Chain object, input and output
 indexes are directly extracted.
 """
 function PredictiveModel(network::Flux.Chain)
-    input_size = size(network[1].weight)[2]
-    output_size = size(network[end].weight)[1]
+    param_layers = [layer for layer in network if has_params(layer)]
+    input_size = size(param_layers[1].weight, 2)
+    output_size = size(param_layers[end].weight, 1)
     input_output_map = [Dict(collect(1:input_size) => collect(1:output_size))]
     return PredictiveModel(
         [deepcopy(network)],

test/runtests.jl

@@ -14,3 +14,4 @@ Random.seed!(123)
 include("utils.jl")
 include("test_predictive_model.jl")
 include("test_newsvendor.jl")
+include("test_gradient.jl")

test/test_gradient.jl

@@ -0,0 +1,41 @@
+# basic model for testing gradient mode
+X = Float32.(ones(1, 1))
+Y = Float32.(ones(1, 1))
+model = ApplicationDrivenLearning.Model()
+@variables(model, begin
+    x >= 0, ApplicationDrivenLearning.Policy
+    d, ApplicationDrivenLearning.Forecast
+end)
+@objective(ApplicationDrivenLearning.Plan(model), Min, x.plan)
+@objective(ApplicationDrivenLearning.Assess(model), Min, x.assess)
+set_optimizer(model, HiGHS.Optimizer)
+set_silent(model)
+ApplicationDrivenLearning.set_forecast_model(model, Chain(Dense(1 => 1)))
+
+@testset "GradientMode Stop Rules" begin
+    # epochs
+    initial_sol = ApplicationDrivenLearning.extract_params(model.forecast)
+    opt = ApplicationDrivenLearning.Options(
+        ApplicationDrivenLearning.GradientMode,
+        epochs = 0,
+    )
+    sol = ApplicationDrivenLearning.train!(model, X, Y, opt)
+    @test initial_sol == sol.params
+
+    # time_limit
+    opt = ApplicationDrivenLearning.Options(
+        ApplicationDrivenLearning.GradientMode,
+        time_limit = 0,
+    )
+    sol = ApplicationDrivenLearning.train!(model, X, Y, opt)
+    @test initial_sol == sol.params
+
+    # gradient norm
+    initial_sol = ApplicationDrivenLearning.extract_params(model.forecast)
+    opt = ApplicationDrivenLearning.Options(
+        ApplicationDrivenLearning.GradientMode,
+        g_tol = Inf,
+    )
+    sol = ApplicationDrivenLearning.train!(model, X, Y, opt)
+    @test initial_sol == sol.params
+end