|
| 1 | +// Licensed to the .NET Foundation under one or more agreements. |
| 2 | +// The .NET Foundation licenses this file to you under the MIT license. |
| 3 | +// See the LICENSE file in the project root for more information. |
| 4 | + |
| 5 | +using Microsoft.ML.Runtime; |
| 6 | +using Microsoft.ML.Runtime.Data; |
| 7 | +using System.Collections.Generic; |
| 8 | +using System.Linq; |
| 9 | + |
| 10 | +namespace Microsoft.ML.Core.Data |
| 11 | +{ |
| 12 | + /// <summary> |
| 13 | + /// A set of 'requirements' to the incoming schema, as well as a set of 'promises' of the outgoing schema. |
| 14 | + /// This is more relaxed than the proper <see cref="ISchema"/>, since it's only a subset of the columns, |
| 15 | + /// and also since it doesn't specify exact <see cref="ColumnType"/>'s for vectors and keys. |
| 16 | + /// </summary> |
| 17 | + public sealed class SchemaShape |
| 18 | + { |
| 19 | + public readonly Column[] Columns; |
| 20 | + |
| 21 | + public sealed class Column |
| 22 | + { |
| 23 | + public enum VectorKind |
| 24 | + { |
| 25 | + Scalar, |
| 26 | + Vector, |
| 27 | + VariableVector |
| 28 | + } |
| 29 | + |
| 30 | + public readonly string Name; |
| 31 | + public readonly VectorKind Kind; |
| 32 | + public readonly DataKind ItemKind; |
| 33 | + public readonly bool IsKey; |
| 34 | + public readonly string[] MetadataKinds; |
| 35 | + |
| 36 | + public Column(string name, VectorKind vecKind, DataKind itemKind, bool isKey, string[] metadataKinds) |
| 37 | + { |
| 38 | + Contracts.CheckNonEmpty(name, nameof(name)); |
| 39 | + Contracts.CheckValue(metadataKinds, nameof(metadataKinds)); |
| 40 | + |
| 41 | + Name = name; |
| 42 | + Kind = vecKind; |
| 43 | + ItemKind = itemKind; |
| 44 | + IsKey = isKey; |
| 45 | + MetadataKinds = metadataKinds; |
| 46 | + } |
| 47 | + } |
| 48 | + |
| 49 | + public SchemaShape(Column[] columns) |
| 50 | + { |
| 51 | + Contracts.CheckValue(columns, nameof(columns)); |
| 52 | + Columns = columns; |
| 53 | + } |
| 54 | + |
| 55 | + /// <summary> |
| 56 | + /// Create a schema shape out of the fully defined schema. |
| 57 | + /// </summary> |
| 58 | + public static SchemaShape Create(ISchema schema) |
| 59 | + { |
| 60 | + Contracts.CheckValue(schema, nameof(schema)); |
| 61 | + var cols = new List<Column>(); |
| 62 | + |
| 63 | + for (int iCol = 0; iCol < schema.ColumnCount; iCol++) |
| 64 | + { |
| 65 | + if (!schema.IsHidden(iCol)) |
| 66 | + { |
| 67 | + Column.VectorKind vecKind; |
| 68 | + var type = schema.GetColumnType(iCol); |
| 69 | + if (type.IsKnownSizeVector) |
| 70 | + vecKind = Column.VectorKind.Vector; |
| 71 | + else if (type.IsVector) |
| 72 | + vecKind = Column.VectorKind.VariableVector; |
| 73 | + else |
| 74 | + vecKind = Column.VectorKind.Scalar; |
| 75 | + |
| 76 | + var kind = type.ItemType.RawKind; |
| 77 | + var isKey = type.ItemType.IsKey; |
| 78 | + |
| 79 | + var metadataNames = schema.GetMetadataTypes(iCol) |
| 80 | + .Select(kvp => kvp.Key) |
| 81 | + .ToArray(); |
| 82 | + cols.Add(new Column(schema.GetColumnName(iCol), vecKind, kind, isKey, metadataNames)); |
| 83 | + } |
| 84 | + } |
| 85 | + return new SchemaShape(cols.ToArray()); |
| 86 | + } |
| 87 | + |
| 88 | + /// <summary> |
| 89 | + /// Returns the column with a specified <paramref name="name"/>, and <c>null</c> if there is no such column. |
| 90 | + /// </summary> |
| 91 | + public Column FindColumn(string name) |
| 92 | + { |
| 93 | + Contracts.CheckValue(name, nameof(name)); |
| 94 | + return Columns.FirstOrDefault(x => x.Name == name); |
| 95 | + } |
| 96 | + |
| 97 | + // REVIEW: I think we should have an IsCompatible method to check if it's OK to use one schema shape |
| 98 | + // as an input to another schema shape. I started writing, but realized that there's more than one way to check for |
| 99 | + // the 'compatibility': as in, 'CAN be compatible' vs. 'WILL be compatible'. |
| 100 | + } |
| 101 | + |
| 102 | + /// <summary> |
| 103 | + /// The 'data reader' takes a certain kind of input and turns it into an <see cref="IDataView"/>. |
| 104 | + /// </summary> |
| 105 | + /// <typeparam name="TSource">The type of input the reader takes.</typeparam> |
| 106 | + public interface IDataReader<in TSource> |
| 107 | + { |
| 108 | + /// <summary> |
| 109 | + /// Produce the data view from the specified input. |
| 110 | + /// Note that <see cref="IDataView"/>'s are lazy, so no actual reading happens here, just schema validation. |
| 111 | + /// </summary> |
| 112 | + IDataView Read(TSource input); |
| 113 | + |
| 114 | + /// <summary> |
| 115 | + /// The output schema of the reader. |
| 116 | + /// </summary> |
| 117 | + ISchema GetOutputSchema(); |
| 118 | + } |
| 119 | + |
| 120 | + /// <summary> |
| 121 | + /// Sometimes we need to 'fit' an <see cref="IDataReader{TIn}"/>. |
| 122 | + /// A DataReader estimator is the object that does it. |
| 123 | + /// </summary> |
| 124 | + public interface IDataReaderEstimator<in TSource, out TReader> |
| 125 | + where TReader : IDataReader<TSource> |
| 126 | + { |
| 127 | + /// <summary> |
| 128 | + /// Train and return a data reader. |
| 129 | + /// |
| 130 | + /// REVIEW: you could consider the transformer to take a different <typeparamref name="TSource"/>, but we don't have such components |
| 131 | + /// yet, so why complicate matters? |
| 132 | + /// </summary> |
| 133 | + TReader Fit(TSource input); |
| 134 | + |
| 135 | + /// <summary> |
| 136 | + /// The 'promise' of the output schema. |
| 137 | + /// It will be used for schema propagation. |
| 138 | + /// </summary> |
| 139 | + SchemaShape GetOutputSchema(); |
| 140 | + } |
| 141 | + |
| 142 | + /// <summary> |
| 143 | + /// The transformer is a component that transforms data. |
| 144 | + /// It also supports 'schema propagation' to answer the question of 'how the data with this schema look after you transform it?'. |
| 145 | + /// </summary> |
| 146 | + public interface ITransformer |
| 147 | + { |
| 148 | + /// <summary> |
| 149 | + /// Schema propagation for transformers. |
| 150 | + /// Returns the output schema of the data, if the input schema is like the one provided. |
| 151 | + /// Returns <c>null</c> iff the schema is invalid (then a call to Transform with this data will fail). |
| 152 | + /// </summary> |
| 153 | + ISchema GetOutputSchema(ISchema inputSchema); |
| 154 | + |
| 155 | + /// <summary> |
| 156 | + /// Take the data in, make transformations, output the data. |
| 157 | + /// Note that <see cref="IDataView"/>'s are lazy, so no actual transformations happen here, just schema validation. |
| 158 | + /// </summary> |
| 159 | + IDataView Transform(IDataView input); |
| 160 | + } |
| 161 | + |
| 162 | + /// <summary> |
| 163 | + /// The estimator (in Spark terminology) is an 'untrained transformer'. It needs to 'fit' on the data to manufacture |
| 164 | + /// a transformer. |
| 165 | + /// It also provides the 'schema propagation' like transformers do, but over <see cref="SchemaShape"/> instead of <see cref="ISchema"/>. |
| 166 | + /// </summary> |
| 167 | + public interface IEstimator<out TTransformer> |
| 168 | + where TTransformer : ITransformer |
| 169 | + { |
| 170 | + /// <summary> |
| 171 | + /// Train and return a transformer. |
| 172 | + /// </summary> |
| 173 | + TTransformer Fit(IDataView input); |
| 174 | + |
| 175 | + /// <summary> |
| 176 | + /// Schema propagation for estimators. |
| 177 | + /// Returns the output schema shape of the estimator, if the input schema shape is like the one provided. |
| 178 | + /// Returns <c>null</c> iff the schema shape is invalid (then a call to <see cref="Fit"/> with this data will fail). |
| 179 | + /// </summary> |
| 180 | + SchemaShape GetOutputSchema(SchemaShape inputSchema); |
| 181 | + } |
| 182 | +} |
0 commit comments