调用trainingPipeline.Fit时遇‘Method not found: Void Module.train()’错误求助

这个错误几乎都是TorchSharp与Microsoft.ML.TorchSharp版本不兼容导致的，结合你的代码细节，给出以下修复方案：

1. 强制统一NuGet包版本 #

卸载当前项目中的Microsoft.ML.TorchSharp和TorchSharp包，重新安装完全相同版本的两个包。ML.NET的TorchSharp绑定依赖特定版本的TorchSharp核心库，版本不匹配会直接引发方法找不到的反射错误。

2. 移除冗余的文本特征处理步骤 #

你的ProcessData方法中手动做了FeaturizeText和Concatenate，但TextClassification Trainer（基于BERT）会自动处理原始文本列的特征提取，手动预处理会和Trainer的内置逻辑冲突。修改后的ProcessData方法如下：

IEstimator<ITransformer> ProcessData() {
    var pipeline = _mlContext.Transforms.Conversion.MapValueToKey(inputColumnName: "Sentiment", outputColumnName: "Label")
        .AppendCacheCheckpoint(_mlContext);
    return pipeline;
}

3. 补充足够的训练数据 #

当前你只使用了1条训练数据，经过TrainTestSplit后训练集数据量极端不足，会导致训练流程异常。至少准备5条以上带标注的样本，比如：

var reviews = new List<ExpenseNote>
{
    new ExpenseNote() { Note = "报销办公用品费用", Sentiment = "Approved"},
    new ExpenseNote() { Note = "超出预算的招待费", Sentiment = "Rejected"},
    new ExpenseNote() { Note = "差旅住宿发票齐全", Sentiment = "Approved"},
    new ExpenseNote() { Note = "无对应合同的采购支出", Sentiment = "Rejected"},
    new ExpenseNote() { Note = "员工加班餐费报销", Sentiment = "Approved"}
};

4. 清理项目缓存并重建 #

删除项目目录下的bin和obj文件夹，然后重新生成项目，避免旧版本的dll缓存干扰。

修改后的完整代码示例 #

using System;
using System.Collections.Generic;
using Microsoft.ML;
using Microsoft.ML.Data;
using Microsoft.ML.TorchSharp;
using TorchSharp;

namespace ConsoleApp5 {
    internal class TextNLPClassifier {
        MLContext _mlContext;
        PredictionEngine<ExpenseNote, IssuePrediction> _predEngine;
        ITransformer _trainedModel;

        public void FitData() {
            _mlContext = new MLContext(seed: 0);

            // 补充足够的训练数据
            var reviews = new List<ExpenseNote>
            {
                new ExpenseNote() { Note = "报销办公用品费用", Sentiment = "Approved"},
                new ExpenseNote() { Note = "超出预算的招待费", Sentiment = "Rejected"},
                new ExpenseNote() { Note = "差旅住宿发票齐全", Sentiment = "Approved"},
                new ExpenseNote() { Note = "无对应合同的采购支出", Sentiment = "Rejected"},
                new ExpenseNote() { Note = "员工加班餐费报销", Sentiment = "Approved"}
            };

            var reviewsDV = _mlContext.Data.LoadFromEnumerable<ExpenseNote>(reviews);

            var pipeline = ProcessData();
            BuildAndTrainModel(reviewsDV, pipeline);

            foreach(var rev in reviews) {
                var review = new ExpenseNote { Note = rev.Note }; // 创建无标注的预测样本
                var prediction = _predEngine.Predict(review);
                Console.WriteLine($"备注: {review.Note} | 预测结果: {prediction.Sentiment}");
            }
        }

        IEstimator<ITransformer> ProcessData() {
            var pipeline = _mlContext.Transforms.Conversion.MapValueToKey(inputColumnName: "Sentiment", outputColumnName: "Label")
                .AppendCacheCheckpoint(_mlContext);
            return pipeline;
        }

        void BuildAndTrainModel(IDataView trainingDataView, IEstimator<ITransformer> pipeline) {
            var trainTestSplit = _mlContext.Data.TrainTestSplit(trainingDataView, testFraction:0.2);
            
            var trainingPipeline = pipeline.Append(_mlContext.MulticlassClassification.Trainers.TextClassification(numberOfClasses: 2, sentence1ColumnName: "Note"))
                .Append(_mlContext.Transforms.Conversion.MapKeyToValue("PredictedLabel"));
            _trainedModel = trainingPipeline.Fit(trainTestSplit.TrainSet);
            _predEngine = _mlContext.Model.CreatePredictionEngine<ExpenseNote, IssuePrediction>(_trainedModel);
        }
    }

    public class ExpenseNote {
        [LoadColumn(0)]
        public string Note { get; set; }
        [LoadColumn(1)]
        public string Sentiment { get; set; }
    }

    public class IssuePrediction {
        [ColumnName("PredictedLabel")]
        public string Sentiment;
    }
}

内容的提问来源于stack exchange，提问作者Josh Reed