fix recommendation Bug (#4774)

dotnet · Feb 4, 2020 · daae78d · daae78d
1 parent 982df42
commit daae78d
Show file tree

Hide file tree

Showing 5 changed files with 49 additions and 25 deletions.
diff --git a/src/Microsoft.ML.CodeGenerator/CodeGenerator/CSharp/CodeGenerator.cs b/src/Microsoft.ML.CodeGenerator/CodeGenerator/CSharp/CodeGenerator.cs
@@ -329,18 +329,11 @@ private string GenerateConsumeModelCSFileContent(string namespaceValue)
         {
             if (_pipeline == null)
                 throw new ArgumentNullException(nameof(_pipeline));
-            try
-            {
-                var node = _pipeline.Nodes.Where(t => t.NodeType == PipelineNodeType.Trainer).First();
-                ITrainerGenerator generator = TrainerGeneratorFactory.GetInstance(node);
-                var trainerString = generator.GenerateTrainer();
-                var trainerUsings = generator.GenerateUsings();
-                return (trainerString, trainerUsings);
-            }
-            catch (Exception)
-            {
-                return (string.Empty, new string[0]);
-            }
+            var node = _pipeline.Nodes.Where(t => t.NodeType == PipelineNodeType.Trainer).First();
+            ITrainerGenerator generator = TrainerGeneratorFactory.GetInstance(node);
+            var trainerString = generator.GenerateTrainer();
+            var trainerUsings = generator.GenerateUsings();
+            return (trainerString, trainerUsings);
         }
 
         internal IList<string> GenerateClassLabels(IDictionary<string, CodeGeneratorSettings.ColumnMapping> columnMapping = default)

diff --git a/src/Microsoft.ML.CodeGenerator/CodeGenerator/CSharp/TrainerGeneratorBase.cs b/src/Microsoft.ML.CodeGenerator/CodeGenerator/CSharp/TrainerGeneratorBase.cs
@@ -97,6 +97,12 @@ private void Initialize(PipelineNode node)
                     value = "\"" + val + "\"";
                 }
 
+                if (type.IsEnum)
+                {
+                    //example: "MatrixFactorizationTrainer.LossFunctionType.SquareLossRegression"
+                    value = $"{type.ReflectedType.Name}.{type.Name}.{kv.Value.ToString()}";
+                }
+
                 if (type == typeof(CustomProperty))
                 {
                     value = kv.Value;

diff --git a/...torTests.Recommendation_GenerateConsoleAppProjectContents_VerifyModelBuilder.approved.txt b/...torTests.Recommendation_GenerateConsoleAppProjectContents_VerifyModelBuilder.approved.txt
@@ -11,6 +11,7 @@ using System.Linq;
 using Microsoft.ML;
 using Microsoft.ML.Data;
 using TestNamespace.Model;
+using Microsoft.ML.Trainers;
 
 namespace TestNamespace.ConsoleApp
 {
@@ -55,9 +56,10 @@ namespace TestNamespace.ConsoleApp
         public static IEstimator<ITransformer> BuildTrainingPipeline(MLContext mlContext)
         {
             // Data process configuration with pipeline data transformations 
-            var dataProcessPipeline = mlContext.Transforms.Concatenate("Out", new[] { "In" });
+            var dataProcessPipeline = mlContext.Transforms.Conversion.MapValueToKey("userId", "userId")
+                                      .Append(mlContext.Transforms.Conversion.MapValueToKey("movieId", "movieId"));
             // Set the training algorithm 
-            var trainer = mlContext.Recommendation().Trainers.MatrixFactorization(labelColumnName: "Label", matrixColumnIndexColumnName: "userId", matrixRowIndexColumnName: "movieId");
+            var trainer = mlContext.Recommendation().Trainers.MatrixFactorization(new MatrixFactorizationTrainer.Options() { MatrixColumnIndexColumnName = "userId", MatrixRowIndexColumnName = "movieId", LabelColumnName = "Label", NumberOfIterations = 10, LearningRate = 0.01f, ApproximationRank = 8, Lambda = 0.01f, LossFunction = MatrixFactorizationTrainer.LossFunctionType.SquareLossRegression, Alpha = 1f, C = 1E-05f });
 
             var trainingPipeline = dataProcessPipeline.Append(trainer);
 

diff --git a/test/Microsoft.ML.CodeGenerator.Tests/ApprovalTests/ConsoleCodeGeneratorTests.cs b/test/Microsoft.ML.CodeGenerator.Tests/ApprovalTests/ConsoleCodeGeneratorTests.cs
@@ -19,6 +19,7 @@
 using Microsoft.ML.CodeGenerator.Templates.Console;
 using Microsoft.ML.CodeGenerator.Utilities;
 using Microsoft.ML.Data;
+using Microsoft.ML.Trainers;
 using Xunit;
 using CodeGenerator = Microsoft.ML.CodeGenerator.CSharp.CodeGenerator;
 
@@ -466,16 +467,30 @@ private CodeGenerator PrepareForRecommendationTask()
             if (mockedPipeline == null)
             {
                 MLContext context = new MLContext();
+                var hyperParam = new Dictionary<string, object>()
+                {
+                    {"MatrixColumnIndexColumnName","userId" },
+                    {"MatrixRowIndexColumnName","movieId" },
+                    {"LabelColumnName","Label" },
+                    {nameof(MatrixFactorizationTrainer.Options.NumberOfIterations), 10 },
+                    {nameof(MatrixFactorizationTrainer.Options.LearningRate), 0.01f },
+                    {nameof(MatrixFactorizationTrainer.Options.ApproximationRank), 8 },
+                    {nameof(MatrixFactorizationTrainer.Options.Lambda), 0.01f },
+                    {nameof(MatrixFactorizationTrainer.Options.LossFunction), MatrixFactorizationTrainer.LossFunctionType.SquareLossRegression },
+                    {nameof(MatrixFactorizationTrainer.Options.Alpha), 1f },
+                    {nameof(MatrixFactorizationTrainer.Options.C), 0.00001f },
+                };
+                var valueToKeyPipelineNode1 = new PipelineNode(nameof(EstimatorName.ValueToKeyMapping), PipelineNodeType.Transform, "userId", "userId");
+                var valueToKeyPipelineNode2 = new PipelineNode(nameof(EstimatorName.ValueToKeyMapping), PipelineNodeType.Transform, "movieId", "movieId");
+                var matrixPipelineNode = new PipelineNode(nameof(TrainerName.MatrixFactorization), PipelineNodeType.Trainer, "Features", "Score", hyperParam);
+                var pipeline = new Pipeline(new PipelineNode[]
+                {
+                    valueToKeyPipelineNode1,
+                    valueToKeyPipelineNode2,
+                    matrixPipelineNode
+                });
 
-                var trainer1 = new SuggestedTrainer(context, new MatrixFactorizationExtension(), new ColumnInformation() {
-                    LabelColumnName = "Label",
-                    UserIdColumnName = "userId",
-                    ItemIdColumnName = "movieId",
-                }, hyperParamSet: null);
-                var transforms1 = new List<SuggestedTransform>() { ColumnConcatenatingExtension.CreateSuggestedTransform(context, new[] { "In" }, "Out") };
-                var inferredPipeline1 = new SuggestedPipeline(transforms1, new List<SuggestedTransform>(), trainer1, context, false);
-
-                mockedPipeline = inferredPipeline1.ToPipeline();
+                mockedPipeline = pipeline;
                 var textLoaderArgs = new TextLoader.Options()
                 {
                     Columns = new[] {

diff --git a/test/Microsoft.ML.CodeGenerator.Tests/TrainerGeneratorTests.cs b/test/Microsoft.ML.CodeGenerator.Tests/TrainerGeneratorTests.cs
@@ -8,6 +8,7 @@
 using Microsoft.ML;
 using Microsoft.ML.AutoML;
 using Microsoft.ML.CodeGenerator.CSharp;
+using Microsoft.ML.Trainers;
 using Xunit;
 
 namespace mlnet.Tests
@@ -295,14 +296,21 @@ public void MatrixFactorizationAdvancedTest()
                 {"MatrixColumnIndexColumnName","userId" },
                 {"MatrixRowIndexColumnName","movieId" },
                 {"LabelColumnName","rating" },
+                {nameof(MatrixFactorizationTrainer.Options.NumberOfIterations), 10 },
+                {nameof(MatrixFactorizationTrainer.Options.LearningRate), 0.01f },
+                {nameof(MatrixFactorizationTrainer.Options.ApproximationRank), 8 },
+                {nameof(MatrixFactorizationTrainer.Options.Lambda), 0.01f },
+                {nameof(MatrixFactorizationTrainer.Options.LossFunction), MatrixFactorizationTrainer.LossFunctionType.SquareLossRegression },
+                {nameof(MatrixFactorizationTrainer.Options.Alpha), 1f },
+                {nameof(MatrixFactorizationTrainer.Options.C), 0.00001f },
             };
             PipelineNode node = new PipelineNode("MatrixFactorization", PipelineNodeType.Trainer, default(string[]), default(string), elementProperties);
             Pipeline pipeline = new Pipeline(new PipelineNode[] { node });
             CodeGenerator codeGenerator = new CodeGenerator(pipeline, null, null);
             var actual = codeGenerator.GenerateTrainerAndUsings();
-            string expectedTrainerString = "MatrixFactorization(matrixColumnIndexColumnName:\"userId\",matrixRowIndexColumnName:\"movieId\",labelColumnName:\"rating\")";
+            string expectedTrainerString = "MatrixFactorization(new MatrixFactorizationTrainer.Options(){MatrixColumnIndexColumnName=\"userId\",MatrixRowIndexColumnName=\"movieId\",LabelColumnName=\"rating\",NumberOfIterations=10,LearningRate=0.01f,ApproximationRank=8,Lambda=0.01f,LossFunction=MatrixFactorizationTrainer.LossFunctionType.SquareLossRegression,Alpha=1f,C=1E-05f})";
             Assert.Equal(expectedTrainerString, actual.Item1);
-            Assert.Null(actual.Item2);
+            Assert.Equal(new string[] { "using Microsoft.ML.Trainers;\r\n" },actual.Item2);
         }
 
         [Fact]