Convert Ensemble stacking SubComponent to IComponentFactory.

eerhardt · eerhardt · commit 77fff0381a53 · 2018-08-15T17:31:18.000-05:00
diff --git a/src/Microsoft.ML.Core/EntryPoints/ComponentFactory.cs b/src/Microsoft.ML.Core/EntryPoints/ComponentFactory.cs
@@ -29,6 +29,25 @@ public interface IComponentFactory<out TComponent>: IComponentFactory
         TComponent CreateComponent(IHostEnvironment env);
     }
 
+    /// <summary>
+    /// A class for creating a component with no extra parameters (other than an <see cref="IHostEnvironment"/>)
+    /// that simply wraps a delegate which creates the component.
+    /// </summary>
+    public class SimpleComponentFactory<TComponent> : IComponentFactory<TComponent>
+    {
+        private Func<IHostEnvironment, TComponent> _factory;
+
+        public SimpleComponentFactory(Func<IHostEnvironment, TComponent> factory)
+        {
+            _factory = factory;
+        }
+
+        public TComponent CreateComponent(IHostEnvironment env)
+        {
+            return _factory(env);
+        }
+    }
+
     /// <summary>
     /// An interface for creating a component when we take one extra parameter (and an <see cref="IHostEnvironment"/>).
     /// </summary>
diff --git a/src/Microsoft.ML.Ensemble/Microsoft.ML.Ensemble.csproj b/src/Microsoft.ML.Ensemble/Microsoft.ML.Ensemble.csproj
@@ -10,6 +10,7 @@
     <ProjectReference Include="..\Microsoft.ML.Data\Microsoft.ML.Data.csproj" />
     <ProjectReference Include="..\Microsoft.ML.StandardLearners\Microsoft.ML.StandardLearners.csproj" />
     <ProjectReference Include="..\Microsoft.ML.Transforms\Microsoft.ML.Transforms.csproj" />
+    <ProjectReference Include="..\Microsoft.ML.FastTree\Microsoft.ML.FastTree.csproj" />
   </ItemGroup>
 
 </Project>
diff --git a/src/Microsoft.ML.Ensemble/OutputCombiners/BaseScalarStacking.cs b/src/Microsoft.ML.Ensemble/OutputCombiners/BaseScalarStacking.cs
@@ -9,7 +9,7 @@
 
 namespace Microsoft.ML.Runtime.Ensemble.OutputCombiners
 {
-    public abstract class BaseScalarStacking<TSigBase> : BaseStacking<Single, TSigBase>
+    public abstract class BaseScalarStacking : BaseStacking<Single>
     {
         internal BaseScalarStacking(IHostEnvironment env, string name, ArgumentsBase args)
             : base(env, name, args)
diff --git a/src/Microsoft.ML.Ensemble/OutputCombiners/BaseStacking.cs b/src/Microsoft.ML.Ensemble/OutputCombiners/BaseStacking.cs
@@ -7,6 +7,7 @@
 using System.Threading.Tasks;
 using Microsoft.ML.Runtime.CommandLine;
 using Microsoft.ML.Runtime.Data;
+using Microsoft.ML.Runtime.EntryPoints;
 using Microsoft.ML.Runtime.Internal.Internallearn;
 using Microsoft.ML.Runtime.Internal.Utilities;
 using Microsoft.ML.Runtime.Model;
@@ -15,7 +16,7 @@
 namespace Microsoft.ML.Runtime.Ensemble.OutputCombiners
 {
     using ColumnRole = RoleMappedSchema.ColumnRole;
-    public abstract class BaseStacking<TOutput, TSigBase> : IStackingTrainer<TOutput>
+    public abstract class BaseStacking<TOutput> : IStackingTrainer<TOutput>
     {
         public abstract class ArgumentsBase
         {
@@ -24,13 +25,10 @@ public abstract class ArgumentsBase
             [TGUI(Label = "Validation Dataset Proportion")]
             public Single ValidationDatasetProportion = 0.3f;
 
-            [Argument(ArgumentType.Multiple, HelpText = "Base predictor for meta learning", ShortName = "bp", SortOrder = 50,
-                Visibility = ArgumentAttribute.VisibilityType.CmdLineOnly)]
-            [TGUI(Label = "Base predictor")]
-            public SubComponent<ITrainer<IPredictorProducing<TOutput>>, TSigBase> BasePredictorType;
+            public abstract IComponentFactory<ITrainer<IPredictorProducing<TOutput>>> BasePredictorFactory { get; set; }
         }
 
-        protected readonly SubComponent<ITrainer<IPredictorProducing<TOutput>>, TSigBase> BasePredictorType;
+        protected readonly IComponentFactory<ITrainer<IPredictorProducing<TOutput>>> BasePredictorFactory;
         protected readonly IHost Host;
         protected IPredictorProducing<TOutput> Meta;
 
@@ -45,10 +43,10 @@ internal BaseStacking(IHostEnvironment env, string name, ArgumentsBase args)
             Host.CheckUserArg(0 <= args.ValidationDatasetProportion && args.ValidationDatasetProportion < 1,
                     nameof(args.ValidationDatasetProportion),
                     "The validation proportion for stacking should be greater than or equal to 0 and less than 1");
-            Host.CheckUserArg(args.BasePredictorType.IsGood(), nameof(args.BasePredictorType));
+            Host.CheckUserArg(args.BasePredictorFactory != null, nameof(args.BasePredictorFactory));
 
             ValidationDatasetProportion = args.ValidationDatasetProportion;
-            BasePredictorType = args.BasePredictorType;
+            BasePredictorFactory = args.BasePredictorFactory;
         }
 
         internal BaseStacking(IHostEnvironment env, string name, ModelLoadContext ctx)
@@ -135,7 +133,7 @@ public void Train(List<FeatureSubsetModel<IPredictorProducing<TOutput>>> models,
             using (var ch = host.Start("Training stacked model"))
             {
                 ch.Check(Meta == null, "Train called multiple times");
-                ch.Check(BasePredictorType != null);
+                ch.Check(BasePredictorFactory != null);
 
                 var maps = new ValueMapper<VBuffer<Single>, TOutput>[models.Count];
                 for (int i = 0; i < maps.Length; i++)
@@ -187,7 +185,7 @@ public void Train(List<FeatureSubsetModel<IPredictorProducing<TOutput>>> models,
                 var view = bldr.GetDataView();
                 var rmd = new RoleMappedData(view, DefaultColumnNames.Label, DefaultColumnNames.Features);
 
-                var trainer = BasePredictorType.CreateInstance(host);
+                var trainer = BasePredictorFactory.CreateComponent(host);
                 if (trainer.Info.NeedNormalization)
                     ch.Warning("The trainer specified for stacking wants normalization, but we do not currently allow this.");
                 Meta = trainer.Train(rmd);
diff --git a/src/Microsoft.ML.Ensemble/OutputCombiners/MultiStacking.cs b/src/Microsoft.ML.Ensemble/OutputCombiners/MultiStacking.cs
@@ -8,7 +8,10 @@
 using Microsoft.ML.Runtime.Data;
 using Microsoft.ML.Runtime.Ensemble.OutputCombiners;
 using Microsoft.ML.Runtime.EntryPoints;
+using Microsoft.ML.Runtime.FastTree;
+using Microsoft.ML.Runtime.Internal.Internallearn;
 using Microsoft.ML.Runtime.Internal.Utilities;
+using Microsoft.ML.Runtime.Learners;
 using Microsoft.ML.Runtime.Model;
 
 [assembly: LoadableClass(typeof(MultiStacking), typeof(MultiStacking.Arguments), typeof(SignatureCombiner),
@@ -20,7 +23,7 @@
 namespace Microsoft.ML.Runtime.Ensemble.OutputCombiners
 {
     using TVectorPredictor = IPredictorProducing<VBuffer<Single>>;
-    public sealed class MultiStacking : BaseStacking<VBuffer<Single>, SignatureMultiClassClassifierTrainer>, ICanSaveModel, IMultiClassOutputCombiner
+    public sealed class MultiStacking : BaseStacking<VBuffer<Single>>, ICanSaveModel, IMultiClassOutputCombiner
     {
         public const string LoadName = "MultiStacking";
         public const string LoaderSignature = "MultiStackingCombiner";
@@ -38,13 +41,28 @@ private static VersionInfo GetVersionInfo()
         [TlcModule.Component(Name = LoadName, FriendlyName = Stacking.UserName)]
         public sealed class Arguments : ArgumentsBase, ISupportMulticlassOutputCombinerFactory
         {
+            [Argument(ArgumentType.Multiple, HelpText = "Base predictor for meta learning", ShortName = "bp", SortOrder = 50,
+                Visibility = ArgumentAttribute.VisibilityType.CmdLineOnly, SignatureType = typeof(SignatureMultiClassClassifierTrainer))]
+            [TGUI(Label = "Base predictor")]
+            public IComponentFactory<ITrainer<IPredictorProducing<VBuffer<Single>>>> BasePredictorType;
+
+            public override IComponentFactory<ITrainer<TVectorPredictor>> BasePredictorFactory
+            {
+                get { return BasePredictorType; }
+                set { BasePredictorType = value; }
+            }
+
             public IMultiClassOutputCombiner CreateComponent(IHostEnvironment env) => new MultiStacking(env, this);
 
             public Arguments()
             {
                 // REVIEW: Perhaps we can have a better non-parametetric learner.
-                BasePredictorType = new SubComponent<ITrainer<TVectorPredictor>, SignatureMultiClassClassifierTrainer>(
-                    "OVA", "p=FastTreeBinaryClassification");
+                BasePredictorType = new SimpleComponentFactory<ITrainer<TVectorPredictor>>(
+                    env => new Ova(env, new Ova.Arguments()
+                    {
+                        PredictorType = new SimpleComponentFactory<ITrainer<IPredictorProducing<Single>>>(
+                            e => new FastTreeBinaryClassificationTrainer(e, new FastTreeBinaryClassificationTrainer.Arguments()))
+                    }));
             }
         }
 
diff --git a/src/Microsoft.ML.Ensemble/OutputCombiners/RegressionStacking.cs b/src/Microsoft.ML.Ensemble/OutputCombiners/RegressionStacking.cs
@@ -7,6 +7,8 @@
 using Microsoft.ML.Runtime.Data;
 using Microsoft.ML.Runtime.Ensemble.OutputCombiners;
 using Microsoft.ML.Runtime.EntryPoints;
+using Microsoft.ML.Runtime.FastTree;
+using Microsoft.ML.Runtime.Internal.Internallearn;
 using Microsoft.ML.Runtime.Model;
 
 [assembly: LoadableClass(typeof(RegressionStacking), typeof(RegressionStacking.Arguments), typeof(SignatureCombiner),
@@ -19,7 +21,7 @@ namespace Microsoft.ML.Runtime.Ensemble.OutputCombiners
 {
     using TScalarPredictor = IPredictorProducing<Single>;
 
-    public sealed class RegressionStacking : BaseScalarStacking<SignatureRegressorTrainer>, IRegressionOutputCombiner, ICanSaveModel
+    public sealed class RegressionStacking : BaseScalarStacking, IRegressionOutputCombiner, ICanSaveModel
     {
         public const string LoadName = "RegressionStacking";
         public const string LoaderSignature = "RegressionStacking";
@@ -37,9 +39,21 @@ private static VersionInfo GetVersionInfo()
         [TlcModule.Component(Name = LoadName, FriendlyName = Stacking.UserName)]
         public sealed class Arguments : ArgumentsBase, ISupportRegressionOutputCombinerFactory
         {
+            [Argument(ArgumentType.Multiple, HelpText = "Base predictor for meta learning", ShortName = "bp", SortOrder = 50,
+                Visibility = ArgumentAttribute.VisibilityType.CmdLineOnly, SignatureType = typeof(SignatureRegressorTrainer))]
+            [TGUI(Label = "Base predictor")]
+            public IComponentFactory<ITrainer<IPredictorProducing<Single>>> BasePredictorType;
+
+            public override IComponentFactory<ITrainer<TScalarPredictor>> BasePredictorFactory
+            {
+                get { return BasePredictorType; }
+                set { BasePredictorType = value; }
+            }
+
             public Arguments()
             {
-                BasePredictorType = new SubComponent<ITrainer<TScalarPredictor>, SignatureRegressorTrainer>("FastTreeRegression");
+                BasePredictorType = new SimpleComponentFactory<ITrainer<TScalarPredictor>>(
+                    env => new FastTreeRegressionTrainer(env, new FastTreeRegressionTrainer.Arguments()));
             }
 
             public IRegressionOutputCombiner CreateComponent(IHostEnvironment env) => new RegressionStacking(env, this);
diff --git a/src/Microsoft.ML.Ensemble/OutputCombiners/Stacking.cs b/src/Microsoft.ML.Ensemble/OutputCombiners/Stacking.cs
@@ -5,9 +5,10 @@
 using System;
 using Microsoft.ML.Runtime;
 using Microsoft.ML.Runtime.CommandLine;
-using Microsoft.ML.Runtime.Data;
 using Microsoft.ML.Runtime.Ensemble.OutputCombiners;
 using Microsoft.ML.Runtime.EntryPoints;
+using Microsoft.ML.Runtime.FastTree;
+using Microsoft.ML.Runtime.Internal.Internallearn;
 using Microsoft.ML.Runtime.Model;
 
 [assembly: LoadableClass(typeof(Stacking), typeof(Stacking.Arguments), typeof(SignatureCombiner), Stacking.UserName, Stacking.LoadName)]
@@ -16,7 +17,7 @@
 namespace Microsoft.ML.Runtime.Ensemble.OutputCombiners
 {
     using TScalarPredictor = IPredictorProducing<Single>;
-    public sealed class Stacking : BaseScalarStacking<SignatureBinaryClassifierTrainer>, IBinaryOutputCombiner, ICanSaveModel
+    public sealed class Stacking : BaseScalarStacking, IBinaryOutputCombiner, ICanSaveModel
     {
         public const string UserName = "Stacking";
         public const string LoadName = "Stacking";
@@ -35,9 +36,21 @@ private static VersionInfo GetVersionInfo()
         [TlcModule.Component(Name = LoadName, FriendlyName = UserName)]
         public sealed class Arguments : ArgumentsBase, ISupportBinaryOutputCombinerFactory
         {
+            [Argument(ArgumentType.Multiple, HelpText = "Base predictor for meta learning", ShortName = "bp", SortOrder = 50,
+                Visibility = ArgumentAttribute.VisibilityType.CmdLineOnly, SignatureType = typeof(SignatureBinaryClassifierTrainer))]
+            [TGUI(Label = "Base predictor")]
+            public IComponentFactory<ITrainer<IPredictorProducing<Single>>> BasePredictorType;
+
+            public override IComponentFactory<ITrainer<TScalarPredictor>> BasePredictorFactory
+            {
+                get { return BasePredictorType; }
+                set { BasePredictorType = value; }
+            }
+
             public Arguments()
             {
-                BasePredictorType = new SubComponent<ITrainer<TScalarPredictor>, SignatureBinaryClassifierTrainer>("FastTreeBinaryClassification");
+                BasePredictorType = new SimpleComponentFactory<ITrainer<TScalarPredictor>>(
+                    env => new FastTreeBinaryClassificationTrainer(env, new FastTreeBinaryClassificationTrainer.Arguments()));
             }
 
             public IBinaryOutputCombiner CreateComponent(IHostEnvironment env) => new Stacking(env, this);

Original file line number	Diff line number	Diff line change
`@@ -9,7 +9,7 @@`
`9`	`9`
`10`	`10`	`namespace Microsoft.ML.Runtime.Ensemble.OutputCombiners`
`11`	`11`	`{`
`12`		`- public abstract class BaseScalarStacking<TSigBase> : BaseStacking<Single, TSigBase>`
	`12`	`+ public abstract class BaseScalarStacking : BaseStacking<Single>`
`13`	`13`	`{`
`14`	`14`	`internal BaseScalarStacking(IHostEnvironment env, string name, ArgumentsBase args)`
`15`	`15`	`: base(env, name, args)`