Model Authoring SDK

The Model Authoring SDK enables you to develop custom machine learning Recipes and Feature Pipelines which can be used in Adobe Experience Platform Data Science Workspace, providing implementable templates in PySpark and Spark (Scala).

This document provides information regarding the various classes found within the Model Authoring SDK.

DataLoader

The DataLoader class encapsulates anything related to the retrieving, filtering, and returning of raw input data. Examples of input data include those for training, scoring, or feature engineering. Data loaders extend the abstract class DataLoader and must override the abstract method load.

PySpark

The following table describes the abstract methods of a PySpark Data Loader class:

Spark

The following table describes the abstract methods of a Spark Data Loader class:

Load data from a Platform dataset

The following example retrieves Platform data by ID and returns a DataFrame, where the dataset ID (datasetId) is a defined property in the configuration file.

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

The DataSaver class encapsulates anything related to storing output data including those from scoring or feature engineering. Data savers extend the abstract class DataSaver and must override the abstract method save.

PySpark

The following table describes the abstract methods of a PySpark Data Saver class:

Spark (Scala)

The following table describes the abstract methods of a Spark Data Saver class:

Save data to a Platform dataset

In order to store data onto a Platform dataset, the properties must be either provided or defined in the configuration file:

• A valid Platform dataset ID to which data will be stored
• The tenant ID belonging to your organization

The following examples store data (prediction) onto a Platform dataset, where the dataset ID (datasetId) and tenant ID (tenantId) are defined properties within the configuration file.

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

The DatasetTransformer class modifies and transforms the structure of a dataset. The Sensei Machine Learning Runtime does not require this component to be defined, and is implemented based on your requirements.

In regards to a feature pipeline, dataset transformers can be used cooporatively with a feature pipeline factory to prepare data for feature engineering.

PySpark

The following table describes the class methods of a PySpark dataset transformer class:

Spark (Scala)

The following table describes the abstract methods of a Spark dataset transformer class:

FeaturePipelineFactory

The FeaturePipelineFactory class contains feature extraction algorithms and defines the stages of a Feature Pipeline from start to finish.

PySpark

The following table describes the class methods of a PySpark FeaturePipelineFactory:

Spark (Scala)

The following table describes the class methods of a Spark FeaturePipelineFactory:

PipelineFactory

The PipelineFactory class encapsulates methods and definitions for model training and scoring, where training logic and algorithms are defined in the form of a Spark Pipeline.

PySpark

The following table describes the class methods of a PySpark PipelineFactory:

Spark (Scala)

The following table describes the class methods of a Spark PipelineFactory:

MLEvaluator

The MLEvaluator class provides methods for defining evaluation metrics and determining training and testing datasets.

PySpark

The following table describes the class methods of a PySpark MLEvaluator:

Spark (Scala)

The following table describes the class methods of a Spark MLEvaluator: