今日表情 ????
Spark-scala 可以使用LightGBM模型，既可以进行分布式训练，也可以进行分布式预测，支持各种参数设置。
支持模型保存，并且保存后的模型和Python等语言是可以相互调用的。
需要注意的是，Spark-scala训练LightGBM模型时， 输入模型的训练数据集需要处理成一个DataFrame，用spark.ml.feature.VectorAssembler将多列特征转换成一个 features向量列，label作为另外一列。
一，环境配置
spark-scala要使用lightgbm模型，pom文件中要配置如下依赖。
<dependency>
<groupId>org.apache.spark</groupId>
<artifactId>spark-mllib_${scala.version}</artifactId>
<version>${spark.version}</version>
<!--spark-ml要去掉pmml-model依赖-->
<exclusions>
<exclusion>
<groupId>org.jpmml</groupId>
<artifactId>pmml-model</artifactId>
</exclusion>
</exclusions>
</dependency>
<dependency>
<groupId>org.jpmml</groupId>
<artifactId>jpmml-sparkml</artifactId>
<version>1.3.4</version>
</dependency>
<dependency>
<groupId>org.jpmml</groupId>
<artifactId>jpmml-lightgbm</artifactId>
<version>1.3.4</version>
</dependency>
二，范例代码
下面我们以二分类问题为例，按照如下几个大家熟悉的步骤进行范例代码演示。
-
1，准备数据
-
2，定义模型
-
3，训练模型
-
4，评估模型
-
5，使用模型
-
6，保存模型
import org.apache.spark.sql.SparkSession
import org.apache.spark.sql.DataFrame
import org.apache.spark.sql.types.{DoubleType, StringType, StructField, StructType, IntegerType}
import org.apache.spark.ml.Pipeline
import org.apache.spark.ml.evaluation.BinaryClassificationEvaluator
import org.apache.spark.ml.evaluation.MulticlassClassificationEvaluator
import org.apache.spark.ml.linalg.Vector
import org.apache.spark.ml.feature.VectorAssembler
import org.apache.spark.ml.attribute.Attribute
import org.apache.spark.ml.feature.{IndexToString, StringIndexer}
import com.microsoft.ml.spark.{lightgbm=>lgb}
import com.google.gson.{JsonObject, JsonParser}
import scala.collection.JavaConverters._
object LgbDemo extends Serializable {
def printlog(info:String): Unit ={
val dt = new java.text.SimpleDateFormat("yyyy-MM-dd HH:mm:ss").format(new java.util.Date)
println("=========="*8+dt)
println(info+"\n")
}
def main(args:Array[String]):Unit= {
/*================================================================================*/
// 一，加载数据
/*================================================================================*/
printlog("step1: preparing data ...")
//加载数据
val spark = SparkSession.builder().getOrCreate()
val dfdata_raw = spark.read.option("header","true")
.option("delimiter", "\t")
.option("inferschema", "true")
.option("nullValue","")
.csv("data/breast_cancer.csv")
dfdata_raw.sample(false,0.1,1).printSchema
//将特征组合成features向量
val feature_cols = dfdata_raw.columns.filter(!Array("label").contains(_))
val cate_cols = Array("mean_radius","mean_texture")
val vectorAssembler = new VectorAssembler().
setInputCols(feature_cols).
setOutputCol("features")
val dfdata = vectorAssembler.transform(dfdata_raw).select("features", "label")
val Array(dftrain,dfval) = dfdata.randomSplit(Array(0.7, .3), 666)
//各个特征的名字存储在了schema 的 metadata中了, 所以可以用特征名指定类别特征
println(dfdata.schema("features").metadata)
dfdata.show(10)
/*================================================================================*/
// 二，定义模型
/*================================================================================*/
printlog("step2: defining model ...")
val lgbclassifier = new lgb.LightGBMClassifier()
.setNumIterations(100)
.setLearningRate(0.1)
.setNumLeaves(31)
.setMinSumHessianInLeaf(0.001)
.setMaxDepth(-1)
.setBoostFromAverage(false)
.setFeatureFraction(1.0)
.setMaxBin(255)
.setLambdaL1(0.0)
.setLambdaL2(0.0)
.setBaggingFraction(1.0)
.setBaggingFreq(0)
.setBaggingSeed(1)
.setBoostingType("gbdt") //rf、dart、goss
.setCategoricalSlotNames(cate_cols)
.setObjective("binary") //binary, multiclass
.setFeaturesCol("features")
.setLabelCol("label")
println(lgbclassifier.explainParams)
/*================================================================================*/
// 三，训练模型
/*================================================================================*/
printlog("step3: training model ...")
val lgbmodel = lgbclassifier.fit(dftrain)
val feature_importances = lgbmodel.getFeatureImportances("gain")
val arr = feature_cols.zip(feature_importances).sortBy[Double](t=> -t._2)
val dfimportance = spark.createDataFrame(arr).toDF("feature_name","feature_importance(gain)")
dfimportance.show(100)
/*================================================================================*/
// 四，评估模型
/*================================================================================*/
printlog("step4: evaluating model ...")
val evaluator = new BinaryClassificationEvaluator()
.setLabelCol("label")
.setRawPredictionCol("rawPrediction")
.setMetricName("areaUnderROC")
val dftrain_result = lgbmodel.transform(dftrain)
val dfval_result = lgbmodel.transform(dfval)
val train_auc = evaluator.evaluate(dftrain_result)
val val_auc = evaluator.evaluate(dfval_result)
println(s"train_auc = ${train_auc}")
println(s"val_auc = ${val_auc}")
/*================================================================================*/
// 五，使用模型
/*================================================================================*/
printlog("step5: using model ...")
//批量预测
val dfpredict = lgbmodel.transform(dfval)
dfpredict.sample(false,0.1,1).show(20)
//对单个样本进行预测
val features = dfval.head().getAs[Vector]("features")
val single_result = lgbmodel.predict(features)
println(single_result)
/*================================================================================*/
// 六，保存模型
/*================================================================================*/
printlog("step6: saving model ...")
//保存到集群,多文件
lgbmodel.write.overwrite().save("lgbmodel.model")
//加载集群模型
println("load model ...")
val lgbmodel_loaded = lgb.LightGBMClassificationModel.load("lgbmodel.model")
val dfresult = lgbmodel_loaded.transform(dfval)
dfresult.show()
//保存到本地,单文件,和Python接口兼容
//lgbmodel.saveNativeModel("lgb_model",true)
//加载本地模型
//val lgbmodel_loaded = LightGBMClassificationModel.loadNativeModelFromFile("lgb_model")
}
}
三，输出参考
运行如上代码之后，可以得到如下输出。
注意 println(lgbclassifier.explainParams)可以获取LightGBM模型各个参数的含义以及默认值。
================================================================================2021-07-17 22:16:29
step1: preparing data ...
root
|-- mean_radius: integer (nullable = true)
|-- mean_texture: integer (nullable = true)
|-- mean_perimeter: double (nullable = true)
|-- mean_area: double (nullable = true)
|-- mean_smoothness: double (nullable = true)
|-- mean_compactness: double (nullable = true)
|-- mean_concavity: double (nullable = true)
|-- mean_concave_points: double (nullable = true)
|-- mean_symmetry: double (nullable = true)
|-- mean_fractal_dimension: double (nullable = true)
|-- radius_error: double (nullable = true)
|-- texture_error: double (nullable = true)
|-- perimeter_error: double (nullable = true)
|-- area_error: double (nullable = true)
|-- smoothness_error: double (nullable = true)
|-- compactness_error: double (nullable = true)
|-- concavity_error: double (nullable = true)
|-- concave_points_error: double (nullable = true)
|-- symmetry_error: double (nullable = true)
|-- fractal_dimension_error: double (nullable = true)
|-- worst_radius: double (nullable = true)
|-- worst_texture: double (nullable = true)
|-- worst_perimeter: double (nullable = true)
|-- worst_area: double (nullable = true)
|-- worst_smoothness: double (nullable = true)
|-- worst_compactness: double (nullable = true)
|-- worst_concavity: double (nullable = true)
|-- worst_concave_points: double (nullable = true)
|-- worst_symmetry: double (nullable = true)
|-- worst_fractal_dimension: double (nullable = true)
|-- label: integer (nullable = true)
{"ml_attr":{"attrs":{"numeric":[{"idx":0,"name":"mean_radius"},{"idx":1,"name":"mean_texture"},{"idx":2,"name":"mean_perimeter"},{"idx":3,"name":"mean_area"},{"idx":4,"name":"mean_smoothness"},{"idx":5,"name":"mean_compactness"},{"idx":6,"name":"mean_concavity"},{"idx":7,"name":"mean_concave_points"},{"idx":8,"name":"mean_symmetry"},{"idx":9,"name":"mean_fractal_dimension"},{"idx":10,"name":"radius_error"},{"idx":11,"name":"texture_error"},{"idx":12,"name":"perimeter_error"},{"idx":13,"name":"area_error"},{"idx":14,"name":"smoothness_error"},{"idx":15,"name":"compactness_error"},{"idx":16,"name":"concavity_error"},{"idx":17,"name":"concave_points_error"},{"idx":18,"name":"symmetry_error"},{"idx":19,"name":"fractal_dimension_error"},{"idx":20,"name":"worst_radius"},{"idx":21,"name":"worst_texture"},{"idx":22,"name":"worst_perimeter"},{"idx":23,"name":"worst_area"},{"idx":24,"name":"worst_smoothness"},{"idx":25,"name":"worst_compactness"},{"idx":26,"name":"worst_concavity"},{"idx":27,"name":"worst_concave_points"},{"idx":28,"name":"worst_symmetry"},{"idx":29,"name":"worst_fractal_dimension"}]},"num_attrs":30}}
+--------------------+-----+
| features|label|
+--------------------+-----+
|[17.0,10.0,122.8,...| 0|
|[20.0,17.0,132.9,...| 0|
|[19.0,21.0,130.0,...| 0|
|[11.0,20.0,77.58,...| 0|
|[20.0,14.0,135.1,...| 0|
|[12.0,15.0,82.57,...| 0|
|[18.0,19.0,119.6,...| 0|
|[13.0,20.0,90.2,5...| 0|
|[13.0,21.0,87.5,5...| 0|
|[12.0,24.0,83.97,...| 0|
+--------------------+-----+
only showing top 10 rows
================================================================================2021-07-17 22:16:29
step2: defining model ...
baggingFraction: Bagging fraction (default: 1.0, current: 1.0)
baggingFreq: Bagging frequency (default: 0, current: 0)
baggingSeed: Bagging seed (default: 3, current: 1)
boostFromAverage: Adjusts initial score to the mean of labels for faster convergence (default: true, current: false)
boostingType: Default gbdt = traditional Gradient Boosting Decision Tree. Options are: gbdt, gbrt, rf (Random Forest), random_forest, dart (Dropouts meet Multiple Additive Regression Trees), goss (Gradient-based One-Side Sampling). (default: gbdt, current: gbdt)
categoricalSlotIndexes: List of categorical column indexes, the slot index in the features column (undefined)
categoricalSlotNames: List of categorical column slot names, the slot name in the features column (current: [Ljava.lang.String;@351fb3fc)
defaultListenPort: The default listen port on executors, used for testing (default: 12400)
earlyStoppingRound: Early stopping round (default: 0)
featureFraction: Feature fraction (default: 1.0, current: 1.0)
featuresCol: features column name (default: features, current: features)
initScoreCol: The name of the initial score column, used for continued training (undefined)
isProvideTrainingMetric: Whether output metric result over training dataset. (default: false)
isUnbalance: Set to true if training data is unbalanced in binary classification scenario (default: false)
labelCol: label column name (default: label, current: label)
lambdaL1: L1 regularization (default: 0.0, current: 0.0)
lambdaL2: L2 regularization (default: 0.0, current: 0.0)
learningRate: Learning rate or shrinkage rate (default: 0.1, current: 0.1)
maxBin: Max bin (default: 255, current: 255)
maxDepth: Max depth (default: -1, current: -1)
minSumHessianInLeaf: Minimal sum hessian in one leaf (default: 0.001, current: 0.001)
modelString: LightGBM model to retrain (default: )
numBatches: If greater than 0, splits data into separate batches during training (default: 0)
numIterations: Number of iterations, LightGBM constructs num_class * num_iterations trees (default: 100, current: 100)
numLeaves: Number of leaves (default: 31, current: 31)
objective: The Objective. For regression applications, this can be: regression_l2, regression_l1, huber, fair, poisson, quantile, mape, gamma or tweedie. For classification applications, this can be: binary, multiclass, or multiclassova. (default: binary, current: binary)
parallelism: Tree learner parallelism, can be set to data_parallel or voting_parallel (default: data_parallel)
predictionCol: prediction column name (default: prediction)
probabilityCol: Column name for predicted class conditional probabilities. Note: Not all models output well-calibrated probability estimates! These probabilities should be treated as confidences, not precise probabilities (default: probability)
rawPredictionCol: raw prediction (a.k.a. confidence) column name (default: rawPrediction)
thresholds: Thresholds in multi-class classification to adjust the probability of predicting each class. Array must have length equal to the number of classes, with values > 0 excepting that at most one value may be 0. The class with largest value p/t is predicted, where p is the original probability of that class and t is the class's threshold (undefined)
timeout: Timeout in seconds (default: 1200.0)
useBarrierExecutionMode: Use new barrier execution mode in Beta testing, off by default. (default: false)
validationIndicatorCol: Indicates whether the row is for training or validation (undefined)
verbosity: Verbosity where lt 0 is Fatal, eq 0 is Error, eq 1 is Info, gt 1 is Debug (default: 1)
weightCol: The name of the weight column (undefined)
================================================================================2021-07-17 22:16:29
step3: training model ...
+--------------------+------------------------+
| feature_name|feature_importance(gain)|
+--------------------+------------------------+
| worst_area| 974.9349449056517|
| worst_perimeter| 885.3691593843923|
|worst_concave_points| 255.67364284247745|
| mean_concave_points| 250.21955942230738|
| worst_texture| 151.07745621304454|
| area_error| 65.75557372416814|
| worst_smoothness| 62.29973236144293|
| mean_smoothness| 19.902610011957194|
| worst_radius| 16.8275272153341|
| mean_area| 12.41261211467938|
| mean_perimeter| 12.127510878875537|
| worst_concavity| 11.414242858900646|
| compactness_error| 10.996194651604892|
| mean_texture| 9.274276675339683|
| concavity_error| 8.009578698471008|
| symmetry_error| 7.93458393366217|
| radius_error| 7.357747321194173|
| worst_symmetry| 5.951699663755868|
|fractal_dimension...| 4.811246624133022|
|concave_points_error| 4.73140145466917|
| worst_compactness| 4.469820723182832|
| texture_error| 4.356178728700959|
| mean_compactness| 3.123736411467967|
| mean_symmetry| 1.9968633063354835|
| mean_concavity| 1.9701941942285224|
| smoothness_error| 1.673042485476758|
|worst_fractal_dim...| 1.3582115541525612|
|mean_fractal_dime...| 0.6050912755332459|
| perimeter_error| 0.3889888676278275|
| mean_radius| 5.684356116234315...|
+--------------------+------------------------+
================================================================================2021-07-17 22:16:30
step4: evaluating model ...
train_auc = 1.0
val_auc = 0.9890340267698758
================================================================================2021-07-17 22:16:31
step5: using model ...
+--------------------+-----+--------------------+--------------------+----------+
| features|label| rawPrediction| probability|prediction|
+--------------------+-----+--------------------+--------------------+----------+
|[9.0,12.0,60.34,2...| 1|[-10.570726382467...|[-9.5707263824679...| 1.0|
|[10.0,16.0,65.85,...| 1|[-10.120435089856...|[-9.1204350898567...| 1.0|
|[10.0,21.0,68.51,...| 1|[-8.8020346337692...|[-7.8020346337692...| 1.0|
|[11.0,14.0,73.53,...| 1|[-10.315758226759...|[-9.3157582267596...| 1.0|
|[11.0,15.0,73.38,...| 1|[-10.086077130817...|[-9.0860771308174...| 1.0|
|[11.0,16.0,74.72,...| 1|[-6.9649803118554...|[-5.9649803118554...| 1.0|
|[11.0,17.0,71.25,...| 1|[-10.694667171248...|[-9.6946671712481...| 1.0|
|[11.0,17.0,75.27,...| 1|[-9.0156792680894...|[-8.0156792680894...| 1.0|
|[11.0,18.0,75.17,...| 1|[-5.7513546284621...|[-4.7513546284621...| 1.0|
|[11.0,18.0,76.38,...| 1|[-4.3134421808792...|[-3.3134421808792...| 1.0|
|[12.0,15.0,82.57,...| 0|[2.49310942805160...|[3.49310942805160...| 0.0|
|[12.0,17.0,78.27,...| 1|[-10.516042459712...|[-9.5160424597122...| 1.0|
|[12.0,18.0,83.19,...| 1|[-9.4899850168431...|[-8.4899850168431...| 1.0|
|[12.0,22.0,78.75,...| 1|[-8.9917629958319...|[-7.9917629958319...| 1.0|
|[14.0,15.0,92.68,...| 1|[-7.2724968676775...|[-6.2724968676775...| 1.0|
|[14.0,15.0,95.77,...| 1|[-5.0143190624015...|[-4.0143190624015...| 1.0|
|[14.0,16.0,96.22,...| 1|[-5.3849620427583...|[-4.3849620427583...| 1.0|
|[14.0,19.0,97.83,...| 1|[-3.3292007261919...|[-2.3292007261919...| 1.0|
|[16.0,14.0,104.3,...| 1|[4.66077729134426...|[5.66077729134426...| 0.0|
|[19.0,24.0,122.0,...| 0|[10.1503565558166...|[11.1503565558166...| 0.0|
+--------------------+-----+--------------------+--------------------+----------+
1.0
================================================================================2021-07-17 22:16:31
step6: saving model ...
load model ...
收工。????
本文Spark-scala 使用 LightGBM 模型训练 二分类模型 代码和数据集，以及训练 多分类模型 和 回归模型 的范例代码和数据集，可以在公众号算法美食屋后台回复关键词 spark+lightgbm 获取。
万水千山总是情，点个在看行不行？????
猜你喜欢❤️
380
扫一扫
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
