Add DataView Type Register (supporting loading in-memory image and other in-memory custom types)

src/Microsoft.ML.Data/Data/SchemaDefinition.cs

@@ -382,6 +382,17 @@ public static SchemaDefinition Create(Type userType, Direction direction = Direc
                 if (memberInfo.GetCustomAttribute<NoColumnAttribute>() != null)
                     continue;
 
+                var customAttributes = memberInfo.GetCustomAttributes();
+                var customTypeAttributes = customAttributes.Where(x => x is DataViewTypeAttribute);
+                if (customTypeAttributes.Count() > 1)
+                    throw Contracts.ExceptParam(nameof(userType), "Member {0} cannot be marked with multiple attributes, {1}, derived from {2}.",
+                        memberInfo.Name, customTypeAttributes, typeof(DataViewTypeAttribute));
+                else if (customTypeAttributes.Count() == 1)
+                {
+                    var customTypeAttribute = (DataViewTypeAttribute)customTypeAttributes.First();
+                    customTypeAttribute.Register();
+                }
+
                 var mappingNameAttr = memberInfo.GetCustomAttribute<ColumnNameAttribute>();
                 string name = mappingNameAttr?.Name ?? memberInfo.Name;
                 // Disallow duplicate names, because the field enumeration order is not actually
@@ -392,37 +403,42 @@ public static SchemaDefinition Create(Type userType, Direction direction = Direc
 
                 InternalSchemaDefinition.GetVectorAndItemType(memberInfo, out bool isVector, out Type dataType);
 
-                PrimitiveDataViewType itemType;
-                var keyAttr = memberInfo.GetCustomAttribute<KeyTypeAttribute>();
-                if (keyAttr != null)
-                {
-                    if (!KeyDataViewType.IsValidDataType(dataType))
-                        throw Contracts.ExceptParam(nameof(userType), "Member {0} marked with KeyType attribute, but does not appear to be a valid kind of data for a key type", memberInfo.Name);
-                    if (keyAttr.KeyCount == null)
-                        itemType = new KeyDataViewType(dataType, dataType.ToMaxInt());
-                    else
-                        itemType = new KeyDataViewType(dataType, keyAttr.KeyCount.Count.GetValueOrDefault());
-                }
-                else
-                    itemType = ColumnTypeExtensions.PrimitiveTypeFromType(dataType);
-
                 // Get the column type.
                 DataViewType columnType;
-                var vectorAttr = memberInfo.GetCustomAttribute<VectorTypeAttribute>();
-                if (vectorAttr != null && !isVector)
-                    throw Contracts.ExceptParam(nameof(userType), $"Member {memberInfo.Name} marked with {nameof(VectorTypeAttribute)}, but does not appear to be a vector type", memberInfo.Name);
-                if (isVector)
+                if (!DataViewTypeManager.Knows(dataType, customAttributes))
                 {
-                    int[] dims = vectorAttr?.Dims;
-                    if (dims != null && dims.Any(d => d < 0))
-                        throw Contracts.ExceptParam(nameof(userType), "Some of member {0}'s dimension lengths are negative");
-                    if (Utils.Size(dims) == 0)
-                        columnType = new VectorDataViewType(itemType, 0);
+                    PrimitiveDataViewType itemType;
+                    var keyAttr = memberInfo.GetCustomAttribute<KeyTypeAttribute>();
+                    if (keyAttr != null)
+                    {
+                        if (!KeyDataViewType.IsValidDataType(dataType))
+                            throw Contracts.ExceptParam(nameof(userType), "Member {0} marked with KeyType attribute, but does not appear to be a valid kind of data for a key type", memberInfo.Name);
+                        if (keyAttr.KeyCount == null)
+                            itemType = new KeyDataViewType(dataType, dataType.ToMaxInt());
+                        else
+                            itemType = new KeyDataViewType(dataType, keyAttr.KeyCount.Count.GetValueOrDefault());
+                    }
+                    else
+                        itemType = ColumnTypeExtensions.PrimitiveTypeFromType(dataType);
+
+                    var vectorAttr = memberInfo.GetCustomAttribute<VectorTypeAttribute>();
+                    if (vectorAttr != null && !isVector)
+                        throw Contracts.ExceptParam(nameof(userType), $"Member {memberInfo.Name} marked with {nameof(VectorTypeAttribute)}, but does not appear to be a vector type", memberInfo.Name);
+                    if (isVector)
+                    {
+                        int[] dims = vectorAttr?.Dims;
+                        if (dims != null && dims.Any(d => d < 0))
+                            throw Contracts.ExceptParam(nameof(userType), "Some of member {0}'s dimension lengths are negative");
+                        if (Utils.Size(dims) == 0)
+                            columnType = new VectorDataViewType(itemType, 0);
+                        else
+                            columnType = new VectorDataViewType(itemType, dims);
+                    }
                     else
-                        columnType = new VectorDataViewType(itemType, dims);
+                        columnType = itemType;
                 }
                 else
-                    columnType = itemType;
+                    columnType = DataViewTypeManager.GetDataViewType(dataType, customAttributes);
 
                 cols.Add(new Column(memberInfo.Name, columnType, name));
             }

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

@@ -269,6 +269,10 @@ private Delegate CreateGetter(DataViewType colType, InternalSchemaDefinition.Col
                         return Utils.MarshalInvoke(delForKey, keyRawType, peek, colType);
                     }
                 }
+                else if (DataViewTypeManager.Knows(colType))
+                {
+                    del = CreateDirectGetterDelegate<int>;
+                }
                 else
                 {
                     // REVIEW: Is this even possible?
@@ -843,7 +847,7 @@ public AnnotationInfo(string kind, T value, DataViewType annotationType = null)
             Contracts.Assert(value != null);
             bool isVector;
             Type itemType;
-            InternalSchemaDefinition.GetVectorAndItemType(typeof(T), "annotation value", out isVector, out itemType);
+            InternalSchemaDefinition.GetVectorAndItemType("annotation value", typeof(T), null, out isVector, out itemType);
 
             if (annotationType == null)
             {

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

@@ -119,7 +119,7 @@ public void AssertRep()
                 Contracts.Assert(Generator.GetMethodInfo().ReturnType == typeof(void));
 
                 // Checks that the return type of the generator is compatible with ColumnType.
-                GetVectorAndItemType(ComputedReturnType, "return type", out bool isVector, out Type itemType);
+                GetVectorAndItemType("return type", ComputedReturnType, null, out bool isVector, out Type itemType);
                 Contracts.Assert(isVector == ColumnType is VectorDataViewType);
                 Contracts.Assert(itemType == ColumnType.GetItemType().RawType);
             }
@@ -147,11 +147,11 @@ public static void GetVectorAndItemType(MemberInfo memberInfo, out bool isVector
             switch (memberInfo)
             {
                 case FieldInfo fieldInfo:
-                    GetVectorAndItemType(fieldInfo.FieldType, fieldInfo.Name, out isVector, out itemType);
+                    GetVectorAndItemType(fieldInfo.Name, fieldInfo.FieldType, fieldInfo.GetCustomAttributes(), out isVector, out itemType);
                     break;
 
                 case PropertyInfo propertyInfo:
-                    GetVectorAndItemType(propertyInfo.PropertyType, propertyInfo.Name, out isVector, out itemType);
+                    GetVectorAndItemType(propertyInfo.Name, propertyInfo.PropertyType, propertyInfo.GetCustomAttributes(), out isVector, out itemType);
                     break;
 
                 default:
@@ -165,13 +165,14 @@ public static void GetVectorAndItemType(MemberInfo memberInfo, out bool isVector
         /// and also the associated data type for this type. If a valid data type could not
         /// be determined, this will throw.
         /// </summary>
-        /// <param name="rawType">The type of the variable to inspect.</param>
         /// <param name="name">The name of the variable to inspect.</param>
+        /// <param name="rawType">The type of the variable to inspect.</param>
+        /// <param name="attributes">Attribute of <paramref name="rawType"/>. It can be <see langword="null"/> if attributes don't exist.</param>
         /// <param name="isVector">Whether this appears to be a vector type.</param>
         /// <param name="itemType">
         /// The corresponding <see cref="PrimitiveDataViewType"/> RawType of the type, or items of this type if vector.
         /// </param>
-        public static void GetVectorAndItemType(Type rawType, string name, out bool isVector, out Type itemType)
+        public static void GetVectorAndItemType(string name, Type rawType, IEnumerable<Attribute> attributes, out bool isVector, out Type itemType)
         {
             // Determine whether this is a vector, and also determine the raw item type.
             isVector = true;
@@ -185,9 +186,12 @@ public static void GetVectorAndItemType(Type rawType, string name, out bool isVe
                 isVector = false;
             }
 
+            // The internal type of string is ReadOnlyMemory<char>. That is, string will be stored as ReadOnlyMemory<char> in IDataView.
             if (itemType == typeof(string))
                 itemType = typeof(ReadOnlyMemory<char>);
-            else if (!itemType.TryGetDataKind(out _))
+            // Check if the itemType extracted from rawType is supported by ML.NET's type system.
+            // It must be one of either ML.NET's pre-defined types or custom types registered by the user.
+            else if (!itemType.TryGetDataKind(out _) && !DataViewTypeManager.Knows(itemType, attributes))
                 throw Contracts.ExceptParam(nameof(rawType), "Could not determine an IDataView type for member {0}", name);
         }
 
@@ -242,7 +246,7 @@ public static InternalSchemaDefinition Create(Type userType, SchemaDefinition us
                     var parameterType = col.ReturnType;
                     if (parameterType == null)
                         throw Contracts.ExceptParam(nameof(userSchemaDefinition), "No return parameter found in computed column.");
-                    GetVectorAndItemType(parameterType, "returnType", out isVector, out dataItemType);
+                    GetVectorAndItemType("returnType", parameterType, null, out isVector, out dataItemType);
                 }
                 // Infer the column name.
                 var colName = string.IsNullOrEmpty(col.ColumnName) ? col.MemberName : col.ColumnName;

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

@@ -319,6 +319,10 @@ private Action<TRow> GenerateSetter(DataViewRow input, int index, InternalSchema
 
                     del = CreateDirectSetter<int>;
                 }
+                else if (DataViewTypeManager.Knows(colType))
+                {
+                    del = CreateDirectSetter<int>;
+                }
                 else
                 {
                     // REVIEW: Is this even possible?

src/Microsoft.ML.Data/Utils/ApiUtils.cs

@@ -3,6 +3,8 @@
 // See the LICENSE file in the project root for more information.
 
 using System;
+using System.Collections.Generic;
+using System.Linq;
 using System.Reflection;
 using System.Reflection.Emit;
 using Microsoft.ML.Data;
@@ -16,14 +18,15 @@ namespace Microsoft.ML
 
     internal static class ApiUtils
     {
-        private static OpCode GetAssignmentOpCode(Type t)
+        private static OpCode GetAssignmentOpCode(Type t, IEnumerable<Attribute> attributes)
         {
             // REVIEW: This should be a Dictionary<Type, OpCode> based solution.
             // DvTypes, strings, arrays, all nullable types, VBuffers and RowId.
             if (t == typeof(ReadOnlyMemory<char>) || t == typeof(string) || t.IsArray ||
                 (t.IsGenericType && t.GetGenericTypeDefinition() == typeof(VBuffer<>)) ||
                 (t.IsGenericType && t.GetGenericTypeDefinition() == typeof(Nullable<>)) ||
-                t == typeof(DateTime) || t == typeof(DateTimeOffset) || t == typeof(TimeSpan) || t == typeof(DataViewRowId))
+                t == typeof(DateTime) || t == typeof(DateTimeOffset) || t == typeof(TimeSpan) ||
+                t == typeof(DataViewRowId) || DataViewTypeManager.Knows(t, attributes))
             {
                 return OpCodes.Stobj;
             }
@@ -56,7 +59,7 @@ internal static Delegate GeneratePeek<TOwn, TRow>(InternalSchemaDefinition.Colum
                 case FieldInfo fieldInfo:
                     Type fieldType = fieldInfo.FieldType;
 
-                    var assignmentOpCode = GetAssignmentOpCode(fieldType);
+                    var assignmentOpCode = GetAssignmentOpCode(fieldType, fieldInfo.GetCustomAttributes());
                     Func<FieldInfo, OpCode, Delegate> func = GeneratePeek<TOwn, TRow, int>;
                     var methInfo = func.GetMethodInfo().GetGenericMethodDefinition()
                         .MakeGenericMethod(typeof(TOwn), typeof(TRow), fieldType);
@@ -65,7 +68,7 @@ internal static Delegate GeneratePeek<TOwn, TRow>(InternalSchemaDefinition.Colum
                 case PropertyInfo propertyInfo:
                     Type propertyType = propertyInfo.PropertyType;
 
-                    var assignmentOpCodeProp = GetAssignmentOpCode(propertyType);
+                    var assignmentOpCodeProp = GetAssignmentOpCode(propertyType, propertyInfo.GetCustomAttributes());
                     Func<PropertyInfo, OpCode, Delegate> funcProp = GeneratePeek<TOwn, TRow, int>;
                     var methInfoProp = funcProp.GetMethodInfo().GetGenericMethodDefinition()
                         .MakeGenericMethod(typeof(TOwn), typeof(TRow), propertyType);
@@ -132,7 +135,7 @@ internal static Delegate GeneratePoke<TOwn, TRow>(InternalSchemaDefinition.Colum
                 case FieldInfo fieldInfo:
                     Type fieldType = fieldInfo.FieldType;
 
-                    var assignmentOpCode = GetAssignmentOpCode(fieldType);
+                    var assignmentOpCode = GetAssignmentOpCode(fieldType, fieldInfo.GetCustomAttributes());
                     Func<FieldInfo, OpCode, Delegate> func = GeneratePoke<TOwn, TRow, int>;
                     var methInfo = func.GetMethodInfo().GetGenericMethodDefinition()
                         .MakeGenericMethod(typeof(TOwn), typeof(TRow), fieldType);
@@ -141,7 +144,7 @@ internal static Delegate GeneratePoke<TOwn, TRow>(InternalSchemaDefinition.Colum
                 case PropertyInfo propertyInfo:
                     Type propertyType = propertyInfo.PropertyType;
 
-                    var assignmentOpCodeProp = GetAssignmentOpCode(propertyType);
+                    var assignmentOpCodeProp = GetAssignmentOpCode(propertyType, propertyInfo.GetCustomAttributes());
                     Func<PropertyInfo, Delegate> funcProp = GeneratePoke<TOwn, TRow, int>;
                     var methInfoProp = funcProp.GetMethodInfo().GetGenericMethodDefinition()
                         .MakeGenericMethod(typeof(TOwn), typeof(TRow), propertyType);

src/Microsoft.ML.DataView/DataViewType.cs

@@ -41,6 +41,10 @@ private protected DataViewType(Type rawType)
         // Object.Equals(Object) and GetHashCode. In classes below where Equals(ColumnType other)
         // is effectively a referencial comparison, there is no need to override base class implementations
         // of Object.Equals(Object) (and GetHashCode) since its also a referencial comparison.
+        /// <summary>
+        /// Return <see langword="true"/> if <see langword="this"/> is equivalent to <paramref name="other"/> and <see langword="false"/> otherwise.
+        /// </summary>
+        /// <param name="other">Another <see cref="DataViewType"/> to be compared with <see langword="this"/>.</param>
         public abstract bool Equals(DataViewType other);
     }
 
@@ -461,4 +465,25 @@ public override bool Equals(DataViewType other)
 
         public override string ToString() => "TimeSpan";
     }
+
+    /// <summary>
+    /// <see cref="DataViewTypeAttribute"/> should be used to decorated class properties and fields, if that class' instances will be loaded as ML.NET <see cref="IDataView"/>.
+    /// The function <see cref="Register"/> will be called to register a <see cref="DataViewType"/> for a <see cref="Type"/> with its <see cref="Attribute"/>s.
+    /// Whenever a value typed to the registered <see cref="Type"/> and its <see cref="Attribute"/>s, that value's type (i.e., a <see cref="DataViewSchema.Column.Type"/>)
+    /// in <see cref="IDataView"/> would be the associated <see cref="DataViewType"/>.
+    /// </summary>
+    [AttributeUsage(AttributeTargets.Field | AttributeTargets.Property, AllowMultiple = false, Inherited = true)]
+    public abstract class DataViewTypeAttribute : Attribute, IEquatable<DataViewTypeAttribute>
+    {
+        /// <summary>
+        /// A function implicitly invoked by ML.NET when processing a custom type. It binds a DataViewType to a custom type plus its attributes.
+        /// </summary>
+        public abstract void Register();
+
+        /// <summary>
+        /// Return <see langword="true"/> if <see langword="this"/> is equivalent to <paramref name="other"/> and <see langword="false"/> otherwise.
+        /// </summary>
+        /// <param name="other">Another <see cref="DataViewTypeAttribute"/> to be compared with <see langword="this"/>.</param>
+        public abstract bool Equals(DataViewTypeAttribute other);
+    }
 }