ML.NET多项式回归中CustomMapping如何设置运行时指定的泛型输入输出类型

泛型化自定义多项式映射实现方案 #

方案1：接口约束泛型参数（编译时类型安全，推荐） #

先定义公共接口约束输入输出类的必要属性，保证转换逻辑可以正常访问所需字段：

// 输入类约束接口
public interface IPolynomialInput
{
    float Level { get; }
    float Salary { get; }
}

// 输出类约束接口
public interface IPolynomialOutput
{
    float[] Features { get; set; }
    float Salary { get; set; }
}

将PolynomialFeatures改造为泛型类，添加接口约束：

[CustomMappingFactoryAttribute("Features")]
public class PolynomialFeatures<TInput, TOutput> : CustomMappingFactory<TInput, TOutput>
    where TInput : class, IPolynomialInput, new()
    where TOutput : class, IPolynomialOutput, new()
{
    private readonly int _degree;
    public PolynomialFeatures(int degree)
    {
        _degree = degree;
    }

    public void Transform(TInput input, TOutput output)
    {
        output.Features = Enumerable.Range(0, _degree + 1).Select(i => (float)(Math.Pow(input.Level, i))).ToArray();
        output.Salary = input.Salary;
    }

    public override Action<TInput, TOutput> GetMapping()
    {
        return Transform;
    }
}

调用方式和原逻辑基本一致，仅需在实例化时指定具体的输入输出类型即可：

// 实例化时指定你自定义的输入、输出类
var polyFeatures = new PolynomialFeatures<你的输入类, 你的输出类>(degree: 2);

var dataProcessPipeline = mLContext.Transforms.CustomMapping(polyFeatures.GetMapping(), contractName: null,inputSchemaDefinition: schema)
                                    .Append(mLContext.Transforms.Concatenate("Features", new[] { "Features" }));

var trainer = mLContext.Regression.Trainers.Ols(featureColumnName: "Features", labelColumnName: "Salary");
var trainingPipeline = dataProcessPipeline.Append(trainer);

方案2：反射实现完全运行时动态类型（无接口约束，灵活度更高） #

如果需要完全不限制输入输出类的结构，可通过反射动态读写属性，无需提前定义接口，还可支持运行时指定要计算的特征列、标签列名称：

[CustomMappingFactoryAttribute("Features")]
public class PolynomialFeatures<TInput, TOutput> : CustomMappingFactory<TInput, TOutput>
    where TInput : class, new()
    where TOutput : class, new()
{
    private readonly int _degree;
    private readonly string _inputFeatureCol;
    private readonly string _labelCol;
    private readonly string _outputFeatureCol;

    public PolynomialFeatures(int degree, string inputFeatureCol = "Level", string labelCol = "Salary", string outputFeatureCol = "Features")
    {
        _degree = degree;
        _inputFeatureCol = inputFeatureCol;
        _labelCol = labelCol;
        _outputFeatureCol = outputFeatureCol;
    }

    public void Transform(TInput input, TOutput output)
    {
        // 反射读取输入属性
        var inputType = typeof(TInput);
        var levelValue = (float)inputType.GetProperty(_inputFeatureCol).GetValue(input);
        var salaryValue = (float)inputType.GetProperty(_labelCol).GetValue(input);

        // 反射写入输出属性
        var outputType = typeof(TOutput);
        var polyFeatures = Enumerable.Range(0, _degree + 1).Select(i => (float)(Math.Pow(levelValue, i))).ToArray();
        outputType.GetProperty(_outputFeatureCol).SetValue(output, polyFeatures);
        outputType.GetProperty(_labelCol).SetValue(output, salaryValue);
    }

    public override Action<TInput, TOutput> GetMapping()
    {
        return Transform;
    }
}

该方案无需输入输出类实现任何接口，仅要求类中存在对应名称的属性即可，适合对灵活度要求高的场景，缺点是反射会带来极小的性能损耗。

注意事项 #

ML.NET的自定义映射要求输入、输出类型必须为公共类，且用到的属性必须具备对应的get/set访问器，否则会出现Schema识别异常。

内容的提问来源于stack exchange，提问作者hussein shaib