public async Task Training()
{
var pipeline = new LearningPipeline();
pipeline.Add(new TextLoader <ManifestDataTraining>(dataPath, useHeader: true, separator: ";"));
pipeline.Add(new ColumnConcatenator("Features",
"Sex",
"Pclass",
"Age",
"Fare",
"SibSp",
"TEmbarked"
));
pipeline.Add(new MissingValuesRowDropper()
{
Column = new[] { "Age" }
});
var classifier = new FastForestBinaryClassifier()
{
NumLeaves = 10, NumTrees = 13, MinDocumentsInLeafs = 5
};
pipeline.Add(classifier);
pipeline.Add(new PredictedLabelColumnOriginalValueConverter()
{
PredictedLabelColumn = "PredictedLabel"
});
model = pipeline.Train <ManifestDataTraining, ManifestPrediction>();
await model.WriteAsync("model.train");
if (model != null)
{
TestModel();
}
}
static async Task Train()
{
//パイプラインの作成
var pipeline = new LearningPipeline();
//訓練データの読み込み
var trainingSets = new TextLoader <TitanicData>(trainSetPath, useHeader: true, separator: ",");
pipeline.Add(trainingSets);
//年齢が欠損値の行を捨てる
pipeline.Add(new MissingValuesRowDropper()
{
Column = new string[] { "Age" }
});
//数値でない変数をOneHotVectorにする
pipeline.Add(new CategoricalOneHotVectorizer("Sex", "Embarked"));
//モデルに使う変数を結合する
pipeline.Add(new ColumnConcatenator("Features",
"Pclass", "Sex", "Age", "SibSp", "Parch", "Fare", "Embarked"));
//交差検証データ
var cvSets = new TextLoader <TitanicData>(crossValidationSetPath, useHeader: true, separator: ",");
//グリッドサーチ用
var n_trees = new int[] { 2, 4, 8, 16, 32, 64, 128 };
var n_leaves = new int[] { 2, 4, 8, 16, 32, 64, 128 };
//最も良い精度
var bestf1 = 0.0;
//最も良い分類器
FastForestBinaryClassifier bestClassifier = null;
//二値分類の評価
var evaluator = new BinaryClassificationEvaluator();
foreach (var nt in n_trees)
{
foreach (var nl in n_leaves)
{
//ランダムフォレストで二値分類
var classifier = new FastForestBinaryClassifier()
{
NumTrees = nt,
NumLeaves = nl
};
pipeline.Add(classifier);
//訓練
var model = pipeline.Train <TitanicData, TitanicPrediction>();
//F1スコア
var metrics = evaluator.Evaluate(model, cvSets);
Console.WriteLine($"#tree = {nt}, #leaf = {nl}, cv_f1={metrics.F1Score}");
if (!double.IsNaN(metrics.F1Score) && metrics.F1Score > bestf1)
{
Console.WriteLine($"[!]Classifier Updated {bestf1} -> {metrics.F1Score} / nt : {nt}, nl : {nl}");
bestf1 = metrics.F1Score;
bestClassifier = classifier;
}
//パイプラインから一旦分類器削除
pipeline.Remove(classifier);
}
}
//グリッドサーチの結果から最も良いモデルを選択してパイプラインに追加
pipeline.Add(bestClassifier);
//訓練
var bestModel = pipeline.Train <TitanicData, TitanicPrediction>();
//訓練誤差
var trainMetrics = evaluator.Evaluate(bestModel, trainingSets);
//交差検証誤差
var cvMetrics = evaluator.Evaluate(bestModel, cvSets);
//テストデータ
var testSets = new TextLoader <TitanicData>(testSetPath, useHeader: true, separator: ",");
//テスト誤差
var testMetrics = evaluator.Evaluate(bestModel, testSets);
//モデルの保存
await bestModel.WriteAsync("model.zip");
//結果表示
Console.WriteLine("### Result ###");
Console.WriteLine("- Selected Classifier");
Console.WriteLine($"NumTree={bestClassifier.NumTrees}, NumLeaves={bestClassifier.NumLeaves}");
Console.WriteLine("- Trian Sets");
Console.WriteLine($"Accuracy={trainMetrics.Accuracy:P2}, " +
$"Precision={trainMetrics.PositivePrecision:P2}, " +
$"Recall={trainMetrics.PositiveRecall:P2}, " +
$"F1score={trainMetrics.F1Score:P2}");
Console.WriteLine("- Cross Validation Sets");
Console.WriteLine($"Accuracy={cvMetrics.Accuracy:P2}, " +
$"Precision={cvMetrics.PositivePrecision:P2}, " +
$"Recall={cvMetrics.PositiveRecall:P2}, " +
$"F1score={cvMetrics.F1Score:P2}");
Console.WriteLine("- Test Sets");
Console.WriteLine($"Accuracy={testMetrics.Accuracy:P2}, " +
$"Precision={testMetrics.PositivePrecision:P2}, " +
$"Recall={testMetrics.PositiveRecall:P2}, " +
$"F1score={testMetrics.F1Score:P2}");
/*
### Result ###
###- Selected Classifier
###NumTree=4, NumLeaves=32
###- Trian Sets
###Accuracy=83.45%, Precision=83.78%, Recall=72.94%, F1score=77.99%
###- Cross Validation Sets
###Accuracy=80.28%, Precision=86.67%, Recall=63.93%, F1score=73.58%
###- Test Sets
###Accuracy=86.58%, Precision=86.79%, Recall=77.97%, F1score=82.14%
*/
}
