模型製作SDK

Model Authoring SDK可讓您開發自訂的機器學習方法和功能管道，這些方法可用於 Adobe Experience Platform Data Science Workspace，提供可實施的範本 PySpark 和 Spark (Scala).

本檔案提供Model Authoring SDK中各種類別的相關資訊。

DataLoader

DataLoader類別會封裝與擷取、篩選和傳回原始輸入資料相關的任何專案。 輸入資料的範例包括訓練、評分或功能工程方面的資料。 資料載入器會擴充抽象類別 DataLoader 和必須覆寫抽象方法 load.

PySpark

下表說明PySpark Data Loader類別的抽象方法：

Spark

下表說明的抽象方法 Spark 資料載入器類別：

從載入資料 Platform 資料集

下列範例會擷取 Platform 依ID分類資料並傳回DataFrame，其中資料集ID (datasetId)是設定檔案中定義的屬性。

PySpark

# PySpark

from sdk.data_loader import DataLoader

class MyDataLoader(DataLoader):
    """
    Implementation of DataLoader which loads a DataFrame and prepares data
    """

    def load_dataset(config_properties, spark, task_id):

        PLATFORM_SDK_PQS_PACKAGE = "com.adobe.platform.query"
        PLATFORM_SDK_PQS_INTERACTIVE = "interactive"

        # prepare variables
        service_token = str(spark.sparkContext.getConf().get("ML_FRAMEWORK_IMS_ML_TOKEN"))
        user_token = str(spark.sparkContext.getConf().get("ML_FRAMEWORK_IMS_TOKEN"))
        org_id = str(spark.sparkContext.getConf().get("ML_FRAMEWORK_IMS_ORG_ID"))
        api_key = str(spark.sparkContext.getConf().get("ML_FRAMEWORK_IMS_CLIENT_ID"))

        dataset_id = str(config_properties.get(task_id))

        # validate variables
        for arg in ['service_token', 'user_token', 'org_id', 'dataset_id', 'api_key']:
            if eval(arg) == 'None':
                raise ValueError("%s is empty" % arg)

        # load dataset through Spark session

        query_options = get_query_options(spark.sparkContext)

        pd = spark.read.format(PLATFORM_SDK_PQS_PACKAGE) \
            .option(query_options.userToken(), user_token) \
            .option(query_options.serviceToken(), service_token) \
            .option(query_options.imsOrg(), org_id) \
            .option(query_options.apiKey(), api_key) \
            .option(query_options.mode(), PLATFORM_SDK_PQS_INTERACTIVE) \
            .option(query_options.datasetId(), dataset_id) \
            .load()
        pd.show()

        # return as DataFrame
        return pd

Spark (Scala)

// Spark

package com.adobe.platform.ml

import java.time.LocalDateTime

import com.adobe.platform.ml.config.ConfigProperties
import com.adobe.platform.query.QSOption
import org.apache.spark.ml.feature.StringIndexer
import org.apache.spark.sql.expressions.Window
import org.apache.spark.sql.functions._
import org.apache.spark.sql.types.{StructType, TimestampType}
import org.apache.spark.sql.{DataFrame, SparkSession}
import org.apache.spark.sql.Column

/**
 * Implementation of DataLoader which loads a DataFrame and prepares data
 */
class MyDataLoader extends DataLoader {

    final val PLATFORM_SDK_PQS_PACKAGE: String = "com.adobe.platform.query"
    final val PLATFORM_SDK_PQS_INTERACTIVE: String = "interactive"
    final val PLATFORM_SDK_PQS_BATCH: String = "batch"

    /**
    *
    * @param configProperties - Configuration Properties map
    * @param sparkSession     - SparkSession
    * @return                 - DataFrame which is loaded for training
    */


  def load_dataset(configProperties: ConfigProperties, sparkSession: SparkSession, taskId: String): DataFrame = {

    require(configProperties != null)
    require(sparkSession != null)

    // Read the configs
    val serviceToken: String = sparkSession.sparkContext.getConf.get("ML_FRAMEWORK_IMS_ML_TOKEN", "").toString
    val userToken: String = sparkSession.sparkContext.getConf.get("ML_FRAMEWORK_IMS_TOKEN", "").toString
    val orgId: String = sparkSession.sparkContext.getConf.get("ML_FRAMEWORK_IMS_ORG_ID", "").toString
    val apiKey: String = sparkSession.sparkContext.getConf.get("ML_FRAMEWORK_IMS_CLIENT_ID", "").toString

    val dataSetId: String = configProperties.get(taskId).getOrElse("")

    // Load the dataset
    var df = sparkSession.read.format(PLATFORM_SDK_PQS_PACKAGE)
      .option(QSOption.userToken, userToken)
      .option(QSOption.serviceToken, serviceToken)
      .option(QSOption.imsOrg, orgId)
      .option(QSOption.apiKey, apiKey)
      .option(QSOption.mode, PLATFORM_SDK_PQS_INTERACTIVE)
      .option(QSOption.datasetId, dataSetId)
      .load()
    df.show()
    df
    }
}

DataSaver

DataSaver類別會封裝與儲存輸出資料相關的任何專案，包括評分或功能工程產生的資料。 資料儲存器會擴充抽象類別 DataSaver 和必須覆寫抽象方法 save.

PySpark

下表說明的抽象方法 PySpark 資料儲存器類別：

Spark (Scala)

下表說明的抽象方法 Spark 資料儲存器類別：

將資料儲存至 Platform 資料集

為了將資料儲存至 Platform 資料集，必須在設定檔案中提供或定義屬性：

• 有效的 Platform 要儲存資料的資料集ID
• 屬於您組織的租使用者ID

以下範例會儲存資料(prediction)至 Platform 資料集，其中資料集ID (datasetId)和租使用者ID (tenantId)的屬性定義於設定檔案中。

PySpark

# PySpark

from sdk.data_saver import DataSaver
from pyspark.sql.types import StringType, TimestampType
from pyspark.sql.functions import col, lit, struct
from .helper import *


class MyDataSaver(DataSaver):
    """
    Implementation of DataSaver which stores a DataFrame to a Platform dataset
    """

    def save(self, config_properties, prediction):

        # Spark context
        sparkContext = prediction._sc

        # preliminary checks
        if config_properties is None:
            raise ValueError("config_properties parameter is null")
        if prediction is None:
            raise ValueError("prediction parameter is null")
        if sparkContext is None:
            raise ValueError("sparkContext parameter is null")

        PLATFORM_SDK_PQS_PACKAGE = "com.adobe.platform.query"

        # prepare variables
        scored_dataset_id = str(config_properties.get("scoringResultsDataSetId"))
        tenant_id = str(config_properties.get("tenant_id"))
        timestamp = "2019-01-01 00:00:00"

        service_token = str(sparkContext.getConf().get("ML_FRAMEWORK_IMS_ML_TOKEN"))
        user_token = str(sparkContext.getConf().get("ML_FRAMEWORK_IMS_TOKEN"))
        org_id = str(sparkContext.getConf().get("ML_FRAMEWORK_IMS_ORG_ID"))
        api_key = str(sparkContext.getConf().get("ML_FRAMEWORK_IMS_CLIENT_ID"))

        # validate variables
       for arg in ['service_token', 'user_token', 'org_id', 'scored_dataset_id', 'api_key', 'tenant_id']:
            if eval(arg) == 'None':
                raise ValueError("%s is empty" % arg)

        scored_df = prediction.withColumn("date", col("date").cast(StringType()))
        scored_df = scored_df.withColumn(tenant_id, struct(col("date"), col("store"), col("prediction")))
        scored_df = scored_df.withColumn("timestamp", lit(timestamp).cast(TimestampType()))
        scored_df = scored_df.withColumn("_id", lit("empty"))
        scored_df = scored_df.withColumn("eventType", lit("empty")

        # store data into dataset

        query_options = get_query_options(sparkContext)

        scored_df.select(tenant_id, "_id", "eventType", "timestamp").write.format(PLATFORM_SDK_PQS_PACKAGE) \
            .option(query_options.userToken(), user_token) \
            .option(query_options.serviceToken(), service_token) \
            .option(query_options.imsOrg(), org_id) \
            .option(query_options.apiKey(), api_key) \
            .option(query_options.datasetId(), scored_dataset_id) \
            .save()

Spark (Scala)

// Spark

package com.adobe.platform.ml

import com.adobe.platform.ml.config.ConfigProperties
import com.adobe.platform.ml.impl.Constants
import com.adobe.platform.ml.sdk.DataSaver
import com.adobe.platform.query.QSOption
import org.apache.spark.sql.DataFrame
import org.apache.spark.sql.functions._
import org.apache.spark.sql.types.TimestampType

/**
 * Implementation of DataSaver which stores a DataFrame to a Platform dataset
 */

class ScoringDataSaver extends DataSaver {

  final val PLATFORM_SDK_PQS_PACKAGE: String = "com.adobe.platform.query"
  final val PLATFORM_SDK_PQS_BATCH: String = "batch"

  /**
    * Method that saves the scoring data into a dataframe
    * @param configProperties  - Configuration Properties map
    * @param dataFrame         - Dataframe with the scoring results
    */

  override def save(configProperties: ConfigProperties, dataFrame: DataFrame): Unit =  {

    require(configProperties != null)
    require(dataFrame != null)

    val predictionColumn = configProperties.get(Constants.PREDICTION_COL).getOrElse(Constants.DEFAULT_PREDICTION)
    val sparkSession = dataFrame.sparkSession

    val serviceToken: String = sparkSession.sparkContext.getConf.get("ML_FRAMEWORK_IMS_ML_TOKEN", "").toString
    val userToken: String = sparkSession.sparkContext.getConf.get("ML_FRAMEWORK_IMS_TOKEN", "").toString
    val orgId: String = sparkSession.sparkContext.getConf.get("ML_FRAMEWORK_IMS_ORG_ID", "").toString
    val apiKey: String = sparkSession.sparkContext.getConf.get("ML_FRAMEWORK_IMS_CLIENT_ID", "").toString
    val tenantId:String = configProperties.get("tenantId").getOrElse("")
    val timestamp:String = "2019-01-01 00:00:00"

    val scoringResultsDataSetId: String = configProperties.get("scoringResultsDataSetId").getOrElse("")
    import sparkSession.implicits._

    var df = dataFrame.withColumn("date", $"date".cast("String"))

    var scored_df  = df.withColumn(tenantId, struct(df("date"), df("store"), df(predictionColumn)))
    scored_df = scored_df.withColumn("timestamp", lit(timestamp).cast(TimestampType))
    scored_df = scored_df.withColumn("_id", lit("empty"))
    scored_df = scored_df.withColumn("eventType", lit("empty"))

    scored_df.select(tenantId, "_id", "eventType", "timestamp").write.format(PLATFORM_SDK_PQS_PACKAGE)
      .option(QSOption.userToken, userToken)
      .option(QSOption.serviceToken, serviceToken)
      .option(QSOption.imsOrg, orgId)
      .option(QSOption.apiKey, apiKey)
      .option(QSOption.datasetId, scoringResultsDataSetId)
      .save()
    }
}

Datasettransformer

DatasetTransformer類別會修改及轉換資料集的結構。 此 Sensei Machine Learning Runtime 不需定義此元件，且會根據您的需求實作。

關於功能管道，資料集轉換器可與功能管道工廠搭配使用，以便為功能工程準備資料。

PySpark

下表說明PySpark資料集轉換器類別的類別方法：

Spark (Scala)

下表說明的抽象方法 Spark 資料集轉換器類別：

FeaturePipelineFactor

FeaturePipelineFactory類別包含功能擷取演演算法，並定義功能管道從開始到結束的階段。

PySpark

下表說明PySpark FeaturePipelineFactory的類別方法：

Spark (Scala)

下表說明的類別方法 Spark FeaturePipelineFactory：

Pipelinefactory

PipelineFactory類別封裝模型訓練和評分的方法和定義，其中訓練邏輯和演演算法的定義形式為 Spark 管道。

PySpark

下表說明PySpark PipelineFactory的類別方法：

Spark (Scala)

下表說明的類別方法 Spark PipelineFactory：

MLEvaluator

MLEvaluator類別提供定義評估量度以及決定訓練和測試資料集的方法。

PySpark

下表說明PySpark MLEvaluator的類別方法：

Spark (Scala)

下表說明的類別方法 Spark MLEvaluator：