.NET ML项目中如何运行时动态生成Feature类？

API 探路者

2026-7-29

解决ML.NET动态特征类的方案

核心结论

ML.NET的PredictionEngine等组件依赖带有ColumnName和LoadColumn特性的静态类型完成数据映射，因此必须保留这些特性。针对动态特征数量的场景，可行方案是运行时动态生成符合要求的静态类，再通过反射调用ML.NET的泛型API。

方案一：用Reflection.Emit动态生成类

这种方式轻量、性能接近静态类，适合大量特征的场景。

1. 动态生成Feature类的代码

using System.Reflection;
using System.Reflection.Emit;
using Microsoft.ML.Data;

public static Type GenerateDynamicFeatureType(IEnumerable<string> featureColumnNames, string labelColumnName)
{
    var assemblyName = new AssemblyName("DynamicMLFeatures");
    var assemblyBuilder = AssemblyBuilder.DefineDynamicAssembly(assemblyName, AssemblyBuilderAccess.Run);
    var moduleBuilder = assemblyBuilder.DefineDynamicModule("DynamicMLFeaturesModule");
    var typeBuilder = moduleBuilder.DefineType("DynamicFeature", TypeAttributes.Public | TypeAttributes.Class);

    // 定义Label属性及特性
    var columnNameAttrCtor = typeof(ColumnNameAttribute).GetConstructor(new[] { typeof(string) });
    var labelProp = typeBuilder.DefineProperty(labelColumnName, PropertyAttributes.None, typeof(float), null);
    var labelField = typeBuilder.DefineField($"_{labelColumnName}", typeof(float), FieldAttributes.Private);
    
    var labelColAttr = new CustomAttributeBuilder(columnNameAttrCtor, new object[] { labelColumnName });
    labelProp.SetCustomAttribute(labelColAttr);

    // Label的Getter/Setter
    var labelGet = typeBuilder.DefineMethod($"get_{labelColumnName}", MethodAttributes.Public | MethodAttributes.SpecialName | MethodAttributes.HideBySig, typeof(float), Type.EmptyTypes);
    var labelGetIl = labelGet.GetILGenerator();
    labelGetIl.Emit(OpCodes.Ldarg_0);
    labelGetIl.Emit(OpCodes.Ldfld, labelField);
    labelGetIl.Emit(OpCodes.Ret);
    
    var labelSet = typeBuilder.DefineMethod($"set_{labelColumnName}", MethodAttributes.Public | MethodAttributes.SpecialName | MethodAttributes.HideBySig, typeof(void), new[] { typeof(float) });
    var labelSetIl = labelSet.GetILGenerator();
    labelSetIl.Emit(OpCodes.Ldarg_0);
    labelSetIl.Emit(OpCodes.Ldarg_1);
    labelSetIl.Emit(OpCodes.Stfld, labelField);
    labelSetIl.Emit(OpCodes.Ret);
    
    labelProp.SetGetMethod(labelGet);
    labelProp.SetSetMethod(labelSet);

    // 定义特征属性及特性
    int columnIndex = 0;
    foreach (var colName in featureColumnNames)
    {
        if (colName.Equals(labelColumnName, StringComparison.OrdinalIgnoreCase))
            continue;

        var prop = typeBuilder.DefineProperty(colName, PropertyAttributes.None, typeof(float), null);
        var field = typeBuilder.DefineField($"_{colName}", typeof(float), FieldAttributes.Private);
        
        // 添加ColumnName和LoadColumn特性
        var loadColAttrCtor = typeof(LoadColumnAttribute).GetConstructor(new[] { typeof(int) });
        var loadAttr = new CustomAttributeBuilder(loadColAttrCtor, new object[] { columnIndex });
        prop.SetCustomAttribute(loadAttr);
        
        var colAttr = new CustomAttributeBuilder(columnNameAttrCtor, new object[] { colName });
        prop.SetCustomAttribute(colAttr);

        // 特征属性的Getter/Setter
        var getMethod = typeBuilder.DefineMethod($"get_{colName}", MethodAttributes.Public | MethodAttributes.SpecialName | MethodAttributes.HideBySig, typeof(float), Type.EmptyTypes);
        var getIl = getMethod.GetILGenerator();
        getIl.Emit(OpCodes.Ldarg_0);
        getIl.Emit(OpCodes.Ldfld, field);
        getIl.Emit(OpCodes.Ret);
        
        var setMethod = typeBuilder.DefineMethod($"set_{colName}", MethodAttributes.Public | MethodAttributes.SpecialName | MethodAttributes.HideBySig, typeof(void), new[] { typeof(float) });
        var setIl = setMethod.GetILGenerator();
        setIl.Emit(OpCodes.Ldarg_0);
        setIl.Emit(OpCodes.Ldarg_1);
        setIl.Emit(OpCodes.Stfld, field);
        setIl.Emit(OpCodes.Ret);
        
        prop.SetGetMethod(getMethod);
        prop.SetSetMethod(setMethod);
        
        columnIndex++;
    }

    return typeBuilder.CreateType()!;
}

2. 使用动态类调用ML.NET API

// 读取CSV并获取列名
var mlContext = new MLContext();
var dataView = mlContext.Data.LoadFromTextFile("data.csv", separatorChar: ',', hasHeader: true);
var columnNames = dataView.Schema.Select(col => col.Name).ToList();
var labelColumnName = "Label";
var featureColumns = columnNames.Where(n => !n.Equals(labelColumnName, StringComparison.OrdinalIgnoreCase)).ToList();

// 生成动态Feature类型
var dynamicFeatureType = GenerateDynamicFeatureType(featureColumns, labelColumnName);

// 反射调用CreatePredictionEngine（假设Prediction类是硬编码的）
var predictionType = typeof(Prediction);
var createPredEngineMethod = typeof(ModelOperationsCatalog)
    .GetMethod(nameof(ModelOperationsCatalog.CreatePredictionEngine))!
    .MakeGenericMethod(dynamicFeatureType, predictionType);

var predictionEngine = createPredEngineMethod.Invoke(mlContext.Model, new[] { trainedModel })!;

3. 创建实例并访问属性

// 创建动态Feature实例
var featureInstance = Activator.CreateInstance(dynamicFeatureType)!;

// 设置属性值（通过反射）
var labelProp = dynamicFeatureType.GetProperty(labelColumnName)!;
labelProp.SetValue(featureInstance, 1.0f);

foreach (var colName in featureColumns)
{
    var prop = dynamicFeatureType.GetProperty(colName)!;
    prop.SetValue(featureInstance, 0.5f); // 替换为实际数据值
}

// 执行预测
var predictMethod = predictionEngine.GetType().GetMethod("Predict")!;
var prediction = predictMethod.Invoke(predictionEngine, new[] { featureInstance })!;

// 获取预测结果
var score = predictionType.GetProperty("Score")!.GetValue(prediction);

方案二：用Roslyn编译动态生成的源代码

如果需要更复杂的类结构（比如不同数据类型的特征），可以用Roslyn动态编译C#代码：

using Microsoft.CodeAnalysis;
using Microsoft.CodeAnalysis.CSharp;
using System.IO;
using System.Linq;
using System.Reflection;
using Microsoft.ML.Data;

public static Type GenerateFeatureTypeWithRoslyn(IEnumerable<string> featureColumnNames, string labelColumnName)
{
    // 生成类的源代码
    var source = $@"
using Microsoft.ML.Data;

public class DynamicFeature
{{
    [ColumnName(""{labelColumnName}"")]
    public float {labelColumnName} {{ get; set; }}
";

    foreach (var colName in featureColumnNames)
    {
        if (colName.Equals(labelColumnName, StringComparison.OrdinalIgnoreCase))
            continue;
        source += $@"
    [ColumnName(""{colName}"")]
    [LoadColumn({featureColumnNames.ToList().IndexOf(colName)})]
    public float {colName} {{ get; set; }}
";
    }

    source += "}";

    // 编译源代码
    var syntaxTree = CSharpSyntaxTree.ParseText(source);
    var references = AppDomain.CurrentDomain.GetAssemblies()
        .Where(a => !a.IsDynamic)
        .Select(a => MetadataReference.CreateFromFile(a.Location))
        .Append(MetadataReference.CreateFromFile(typeof(ColumnNameAttribute).Assembly.Location));

    var compilation = CSharpCompilation.Create("DynamicMLFeatures")
        .WithOptions(new CSharpCompilationOptions(OutputKind.DynamicallyLinkedLibrary))
        .AddReferences(references)
        .AddSyntaxTrees(syntaxTree);

    using var ms = new MemoryStream();
    var result = compilation.Emit(ms);

    if (!result.Success)
    {
        var errors = result.Diagnostics.Where(d => d.Severity == DiagnosticSeverity.Error);
        throw new InvalidOperationException($"编译失败: {string.Join("\n", errors.Select(e => e.GetMessage()))}");
    }

    ms.Seek(0, SeekOrigin.Begin);
    var assembly = Assembly.Load(ms.ToArray());
    return assembly.GetType("DynamicFeature")!;
}

关键注意事项

特性必须保留：ColumnName用于关联CSV列名和类属性，LoadColumn用于指定列索引，ML.NET的核心组件依赖这些特性完成数据绑定。
数据类型匹配：如果CSV中特征的类型不是float，需要从IDataView.Schema中获取列的实际类型，调整动态类的属性类型（比如double、int）。
泛型方法调用：由于动态类型无法在编译时作为泛型参数，必须通过反射的MakeGenericMethod来调用CreatePredictionEngine等泛型API。

内容的提问来源于stack exchange，提问作者Thomas Bont

.NET ML项目中如何运行时动态生成Feature类？

核心结论 #

方案一：用Reflection.Emit动态生成类 #