使用ML.NET实现猜动画片台词

前面几篇主要内容出自微软官方，经我特意修改的案例的文章：

使用ML.NET实现情感分析[新手篇]

使用ML.NET预测纽约出租车费

.NET Core玩转机器学习

使用ML.NET实现情感分析[新手篇]后补

相信看过后大家对ML.NET有了一定的了解了，由于目前还是0.1的版本，也没有更多官方示例放出来，大家普遍觉得提供的特性还不够强大，所以处在观望状态也是能理解的。

本文结合Azure提供的语音识别服务，向大家展示另一种ML.NET有趣的玩法——猜动画片台词。

这个场景特别容易想像，是一种你说我猜的游戏，我会事先用ML.NET对若干动画片的台词进行分类学习，然后使用麦克风，让使用者随便说一句动画片的台词（当然得是数据集中已存在的，没有的不要搞事情呀！），然后来预测出自哪一部。跟随我动手做做看。

准备工作

---

这次需要使用Azure的认知服务中一项API——Speaker Recognition，目前还处于免费试用阶段，打开https://azure.microsoft.com/zh-cn/try/cognitive-services/?api=speaker-recognition，能看到如下页面：

点击获取API密钥，用自己的Azure账号登录，然后就能看到自己的密钥了，类似如下图：

 

创建项目

---

这一次请注意，我们要创建一个.NET Framework 4.6.1或以上版本的控制台应用程序，通过NuGet分别引用三个类库：Microsoft.ML，JiebaNet.Analyser，Microsoft.CognitiveServices.Speech。

然后把编译平台修改成x64，而不是Any CPU。（这一点非常重要）

 

代码分解

---

在Main函数部分，我们只需要关心几个主要步骤，先切词，然后训练模型，最后在一个循环中等待使用者说话，用模型进行预测。

static void Main(string[] args)
{
    Segment(_dataPath, _dataTrainPath);
    var model = Train();
    Evaluate(model);
    ConsoleKeyInfo x;
    do
    {
        var speech = Recognize();
        speech.Wait();
        Predict(model, speech.Result);
        Console.WriteLine("\nRecognition done. Your Choice (0: Stop Any key to continue): ");
        x = Console.ReadKey(true);
    } while (x.Key != ConsoleKey.D0);
}

初始化的变量主要就是训练数据，Azure语音识别密钥等。注意YourServiceRegion的值是“westus”，而不是网址。

const string SubscriptionKey = "你的密钥";
const string YourServiceRegion = "westus";
const string _dataPath = @".\data\dubs.txt";
const string _dataTrainPath = @".\data\dubs_result.txt";

定义数据结构和预测结构和我之前的文章一样，没有什么特别之处。

public class DubbingData
{
    [Column(ordinal: "0")]
    public string DubbingText;
    [Column(ordinal: "1", name: "Label")]
    public string Label;
}

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

 切记部分注意对分隔符的过滤。

public static void Segment(string source, string result)
{
    var segmenter = new JiebaSegmenter();
    using (var reader = new StreamReader(source))
    {
        using (var writer = new StreamWriter(result))
        {
            while (true)
            {
                var line = reader.ReadLine();
                if (string.IsNullOrWhiteSpace(line))
                    break;
                var parts = line.Split(new[] { '\t' }, StringSplitOptions.RemoveEmptyEntries);
                if (parts.Length != 2) continue;
                var segments = segmenter.Cut(parts[0]);
                writer.WriteLine("{0}\t{1}", string.Join(" ", segments), parts[1]);
            }
        }
    }
}

训练部分依然使用熟悉的多分类训练器StochasticDualCoordinateAscentClassifier。TextFeaturizer用于对文本内容向量化处理。

public static PredictionModel<DubbingData, DubbingPrediction> Train()
{
    var pipeline = new LearningPipeline();
    pipeline.Add(new TextLoader<DubbingData>(_dataTrainPath, useHeader: false, separator: "tab"));
    pipeline.Add(new TextFeaturizer("Features", "DubbingText"));
    pipeline.Add(new Dictionarizer("Label"));
    pipeline.Add(new StochasticDualCoordinateAscentClassifier());
    pipeline.Add(new PredictedLabelColumnOriginalValueConverter() { PredictedLabelColumn = "PredictedLabel" });
    var model = pipeline.Train<DubbingData, DubbingPrediction>();
    return model;
}

验证部分这次重点是看损失程度分数。

public static void Evaluate(PredictionModel<DubbingData, DubbingPrediction> model)
{
    var testData = new TextLoader<DubbingData>(_dataTrainPath, useHeader: false, separator: "tab");
    var evaluator = new ClassificationEvaluator();
    var metrics = evaluator.Evaluate(model, testData);
    Console.WriteLine();
    Console.WriteLine("PredictionModel quality metrics evaluation");
    Console.WriteLine("------------------------------------------");
    //Console.WriteLine($"TopKAccuracy: {metrics.TopKAccuracy:P2}");
    Console.WriteLine($"LogLoss: {metrics.LogLoss:P2}");
}

预测部分没有什么大变化，就是对中文交互进行了友好展示。

public static void Predict(PredictionModel<DubbingData, DubbingPrediction> model, string sentence)
{
    IEnumerable<DubbingData> sentences = new[]
    {
        new DubbingData
        {
            DubbingText = sentence
        }
    };

    var segmenter = new JiebaSegmenter();
    foreach (var item in sentences)
    {
        item.DubbingText = string.Join(" ", segmenter.Cut(item.DubbingText));
    }

    IEnumerable<DubbingPrediction> predictions = model.Predict(sentences);
    Console.WriteLine();
    Console.WriteLine("Category Predictions");
    Console.WriteLine("---------------------");

    var sentencesAndPredictions = sentences.Zip(predictions, (sentiment, prediction) => (sentiment, prediction));
    foreach (var item in sentencesAndPredictions)
    {
        Console.WriteLine($"台词: {item.sentiment.DubbingText.Replace(" ", string.Empty)} | 来自动画片: {item.prediction.PredictedLabel}");
    }
    Console.WriteLine();
}

Azure语音识别的调用如下。

static async Task<string> Recognize()
{
    var factory = SpeechFactory.FromSubscription(SubscriptionKey, YourServiceRegion);
    var lang = "zh-cn";

    using (var recognizer = factory.CreateSpeechRecognizer(lang))
    {
        Console.WriteLine("Say something...");

        var result = await recognizer.RecognizeAsync().ConfigureAwait(false);

        if (result.RecognitionStatus != RecognitionStatus.Recognized)
        {
            Console.WriteLine($"There was an error. Status:{result.RecognitionStatus.ToString()}, Reason:{result.RecognitionFailureReason}");
            return null;
        }
        else
        {
            Console.WriteLine($"We recognized: {result.RecognizedText}");
            return result.RecognizedText;
        }
    }
}

运行过程如下：

虽然这看上去有点幼稚，不过一样让你开心一笑了，不是么？请期待更多有趣的案例。

本文使用的数据集：下载

完整的代码如下：

using System;
using Microsoft.ML.Models;
using Microsoft.ML.Runtime;
using Microsoft.ML.Runtime.Api;
using Microsoft.ML.Trainers;
using Microsoft.ML.Transforms;
using System.Collections.Generic;
using System.Linq;
using Microsoft.ML;
using JiebaNet.Segmenter;
using System.IO;
using Microsoft.CognitiveServices.Speech;
using System.Threading.Tasks;

namespace DubbingRecognition
{
    class Program
    {
        public class DubbingData
        {
            [Column(ordinal: "0")]
            public string DubbingText;
            [Column(ordinal: "1", name: "Label")]
            public string Label;
        }

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

        const string SubscriptionKey = "你的密钥";
        const string YourServiceRegion = "westus";
        const string _dataPath = @".\data\dubs.txt";
        const string _dataTrainPath = @".\data\dubs_result.txt";


        static void Main(string[] args)
        {
            Segment(_dataPath, _dataTrainPath);
            var model = Train();
            Evaluate(model);
            ConsoleKeyInfo x;
            do
            {
                var speech = Recognize();
                speech.Wait();
                Predict(model, speech.Result);
                Console.WriteLine("\nRecognition done. Your Choice (0: Stop Any key to continue): ");
                x = Console.ReadKey(true);
            } while (x.Key != ConsoleKey.D0);
        }

        public static void Segment(string source, string result)
        {
            var segmenter = new JiebaSegmenter();
            using (var reader = new StreamReader(source))
            {
                using (var writer = new StreamWriter(result))
                {
                    while (true)
                    {
                        var line = reader.ReadLine();
                        if (string.IsNullOrWhiteSpace(line))
                            break;
                        var parts = line.Split(new[] { '\t' }, StringSplitOptions.RemoveEmptyEntries);
                        if (parts.Length != 2) continue;
                        var segments = segmenter.Cut(parts[0]);
                        writer.WriteLine("{0}\t{1}", string.Join(" ", segments), parts[1]);
                    }
                }
            }
        }

        public static PredictionModel<DubbingData, DubbingPrediction> Train()
        {
            var pipeline = new LearningPipeline();
            pipeline.Add(new TextLoader<DubbingData>(_dataTrainPath, useHeader: false, separator: "tab"));

            //pipeline.Add(new ColumnConcatenator("Features", "DubbingText"));

            pipeline.Add(new TextFeaturizer("Features", "DubbingText"));
            //pipeline.Add(new TextFeaturizer("Label", "Category"));
            pipeline.Add(new Dictionarizer("Label"));
            pipeline.Add(new StochasticDualCoordinateAscentClassifier());
            pipeline.Add(new PredictedLabelColumnOriginalValueConverter() { PredictedLabelColumn = "PredictedLabel" });
            var model = pipeline.Train<DubbingData, DubbingPrediction>();
            return model;
        }

        public static void Evaluate(PredictionModel<DubbingData, DubbingPrediction> model)
        {
            var testData = new TextLoader<DubbingData>(_dataTrainPath, useHeader: false, separator: "tab");
            var evaluator = new ClassificationEvaluator();
            var metrics = evaluator.Evaluate(model, testData);
            Console.WriteLine();
            Console.WriteLine("PredictionModel quality metrics evaluation");
            Console.WriteLine("------------------------------------------");
            //Console.WriteLine($"TopKAccuracy: {metrics.TopKAccuracy:P2}");
            Console.WriteLine($"LogLoss: {metrics.LogLoss:P2}");
        }

        public static void Predict(PredictionModel<DubbingData, DubbingPrediction> model, string sentence)
        {
            IEnumerable<DubbingData> sentences = new[]
            {
                new DubbingData
                {
                    DubbingText = sentence
                }
            };

            var segmenter = new JiebaSegmenter();
            foreach (var item in sentences)
            {
                item.DubbingText = string.Join(" ", segmenter.Cut(item.DubbingText));
            }

            IEnumerable<DubbingPrediction> predictions = model.Predict(sentences);
            Console.WriteLine();
            Console.WriteLine("Category Predictions");
            Console.WriteLine("---------------------");

            var sentencesAndPredictions = sentences.Zip(predictions, (sentiment, prediction) => (sentiment, prediction));
            foreach (var item in sentencesAndPredictions)
            {
                Console.WriteLine($"台词: {item.sentiment.DubbingText.Replace(" ", string.Empty)} | 来自动画片: {item.prediction.PredictedLabel}");
            }
            Console.WriteLine();
        }
        static async Task<string> Recognize()
        {
            var factory = SpeechFactory.FromSubscription(SubscriptionKey, YourServiceRegion);
            var lang = "zh-cn";

            using (var recognizer = factory.CreateSpeechRecognizer(lang))
            {
                Console.WriteLine("Say something...");

                var result = await recognizer.RecognizeAsync().ConfigureAwait(false);

                if (result.RecognitionStatus != RecognitionStatus.Recognized)
                {
                    Console.WriteLine($"There was an error. Status:{result.RecognitionStatus.ToString()}, Reason:{result.RecognitionFailureReason}");
                    return null;
                }
                else
                {
                    Console.WriteLine($"We recognized: {result.RecognizedText}");
                    return result.RecognizedText;
                }
            }
        }
    }
}