Issue 3234: use model schema type instead of class definition schema (#5228)

* fix issue 3234, use model schema type instead of class definition schema

* refine comments

* change test to test on both case

Author: frank-dong-ms-zz

src/Microsoft.ML.Data/DataView/DataViewConstructionUtils.cs

@@ -71,6 +71,13 @@ internal static SchemaDefinition GetSchemaDefinition<TRow>(IHostEnvironment env,
                 var schemaDefinitionCol = schemaDefinition.FirstOrDefault(c => c.ColumnName == name);
                 if (schemaDefinitionCol == null)
                     throw env.Except($"Type should contain a member named {name}");
+
+                //Always use column type from model as this type can be more specific.
+                //This can be corner case:
+                //For example, we can load an model whose schema contains Vector<Single, 38>
+                //and define this field in input class as float[] without specific array length.
+                schemaDefinitionCol.ColumnType = col.Type;
+
                 var annotations = col.Annotations;
                 if (annotations != null)
                 {

test/Microsoft.ML.Functional.Tests/ModelFiles.cs

@@ -217,6 +217,7 @@ public class ModelInput
             public string[] CategoricalFeatures;
             public float[] NumericalFeatures;
 #pragma warning restore SA1401
+            public float Label;
         }
 
         public class ModelOutput
@@ -233,14 +234,29 @@ public void LoadModelWithOptionalColumnTransform()
             SchemaDefinition inputSchemaDefinition = SchemaDefinition.Create(typeof(ModelInput));
             inputSchemaDefinition[nameof(ModelInput.CategoricalFeatures)].ColumnType = new VectorDataViewType(TextDataViewType.Instance, 5);
             inputSchemaDefinition[nameof(ModelInput.NumericalFeatures)].ColumnType = new VectorDataViewType(NumberDataViewType.Single, 3);
+
             var mlContext = new MLContext(1);
             ITransformer trainedModel;
             DataViewSchema dataViewSchema;
             trainedModel = mlContext.Model.Load(TestCommon.GetDataPath(DataDir, "backcompat", "modelwithoptionalcolumntransform.zip"), out dataViewSchema);
+
+            var modelInput = new ModelInput()
+            {
+                CategoricalFeatures = new[] { "ABC", "ABC", "ABC", "ABC", "ABC" },
+                NumericalFeatures = new float[] { 1, 1, 1 },
+                Label = 1
+            };
+
+            // test create prediction engine with user defined schema
             var model = mlContext.Model.CreatePredictionEngine<ModelInput, ModelOutput>(trainedModel, inputSchemaDefinition: inputSchemaDefinition);
-            var prediction = model.Predict(new ModelInput() { CategoricalFeatures = new[] { "ABC", "ABC", "ABC", "ABC", "ABC" }, NumericalFeatures = new float [] { 1, 1, 1 } });
+            var prediction = model.Predict(modelInput);
+
+            // test create prediction engine with schema loaded from model
+            var model2 = mlContext.Model.CreatePredictionEngine<ModelInput, ModelOutput>(trainedModel, inputSchema: dataViewSchema);
+            var prediction2 = model2.Predict(modelInput);
 
             Assert.Equal(1, prediction.Score[0]);
+            Assert.Equal(1, prediction2.Score[0]);
         }
 
         [Fact]