使用模型创作SDK创建特征管线

dataset_id = str(config_properties.get(dataset_id))

# PySpark

from pyspark.sql.types import StringType, TimestampType
from pyspark.sql.functions import col, lit, struct
import logging

class MyDataLoader(DataLoader):
    def load_dataset(config_properties, spark, tenant_id, dataset_id):
    PLATFORM_SDK_PQS_PACKAGE = "com.adobe.platform.query"
    PLATFORM_SDK_PQS_INTERACTIVE = "interactive"

    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(dataset_id))

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

    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()

    # Get the distinct values of the dataframe
    pd = pd.distinct()

    # Flatten the data
    if tenant_id in pd.columns:
        pd = pd.select(col(tenant_id + ".*"))

    return pd

# PySpark

from sdk.dataset_transformer import DatasetTransformer
from pyspark.ml.feature import StringIndexer
from pyspark.sql.types import IntegerType
from pyspark.sql.functions import unix_timestamp, from_unixtime, to_date, lit, lag, udf, date_format, lower, col, split, explode
from pyspark.sql import Window
from .helper import setupLogger

class MyDatasetTransformer(DatasetTransformer):
    logger = setupLogger(__name__)

    def transform(self, config_properties, dataset):
        tenant_id = str(config_properties.get("tenantId"))

        # Flatten the data
        if tenant_id in dataset.columns:
            self.logger.info("Flatten the data before transformation")
            dataset = dataset.select(col(tenant_id + ".*"))
            dataset.show()

        # Convert isHoliday boolean value to Int
        # Rename the column to holiday and drop isHoliday
        pd = dataset.withColumn("holiday", col("isHoliday").cast(IntegerType())).drop("isHoliday")
        pd.show()

        # Get the week and year from date
        pd = pd.withColumn("week", date_format(to_date("date", "MM/dd/yy"), "w").cast(IntegerType()))
        pd = pd.withColumn("year", date_format(to_date("date", "MM/dd/yy"), "Y").cast(IntegerType()))

        # Convert the date to TimestampType
        pd = pd.withColumn("date", to_date(unix_timestamp(pd["date"], "MM/dd/yy").cast("timestamp")))

        # Convert categorical data
        indexer = StringIndexer(inputCol="storeType", outputCol="storeTypeIndex")
        pd = indexer.fit(pd).transform(pd)

        # Get the WeeklySalesAhead and WeeklySalesLag column values
        window = Window.orderBy("date").partitionBy("store")
        pd = pd.withColumn("weeklySalesLag", lag("weeklySales", 1).over(window)).na.drop(subset=["weeklySalesLag"])
        pd = pd.withColumn("weeklySalesAhead", lag("weeklySales", -1).over(window)).na.drop(subset=["weeklySalesAhead"])
        pd = pd.withColumn("weeklySalesScaled", lag("weeklySalesAhead", -1).over(window)).na.drop(subset=["weeklySalesScaled"])
        pd = pd.withColumn("weeklySalesDiff", (pd['weeklySales'] - pd['weeklySalesLag'])/pd['weeklySalesLag'])

        pd = pd.na.drop()
        self.logger.debug("Transformed dataset count is %s " % pd.count())

        # return transformed dataframe
        return pd

# PySpark

from pyspark.ml import Pipeline
from pyspark.ml.regression import GBTRegressor
from pyspark.ml.feature import VectorAssembler

import numpy as np

from sdk.pipeline_factory import PipelineFactory

class MyFeaturePipelineFactory(FeaturePipelineFactory):

    def apply(self, config_properties):
        if config_properties is None:
            raise ValueError("config_properties parameter is null")

        tenant_id = str(config_properties.get("tenantId"))
        input_features = str(config_properties.get("ACP_DSW_INPUT_FEATURES"))

        if input_features is None:
            raise ValueError("input_features parameter is null")
        if input_features.startswith(tenant_id):
            input_features = input_features.replace(tenant_id + ".", "")

        learning_rate = float(config_properties.get("learning_rate"))
        n_estimators = int(config_properties.get("n_estimators"))
        max_depth = int(config_properties.get("max_depth"))

        feature_list = list(input_features.split(","))
        feature_list.remove("date")
        feature_list.remove("storeType")

        cols = np.array(feature_list)

        # Gradient-boosted tree estimator
        gbt = GBTRegressor(featuresCol='features', labelCol='weeklySalesAhead', predictionCol='prediction',
                       maxDepth=max_depth, maxBins=n_estimators, stepSize=learning_rate)

        # Assemble the fields to a vector
        assembler = VectorAssembler(inputCols=cols, outputCol="features")

        # Construct the pipeline
        pipeline = Pipeline(stages=[assembler, gbt])

        return pipeline

    def train(self, config_properties, dataframe):
        pass

    def score(self, config_properties, dataframe, model):
        pass

    def getParamMap(self, config_properties, sparkSession):
        return None

# PySpark

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


class MyDataSaver(DataSaver):
    def save(self, configProperties, data_feature):

        # Spark context
        sparkContext = data_features._sc

        # preliminary checks
        if configProperties is None:
            raise ValueError("configProperties parameter is null")
        if data_features is None:
            raise ValueError("data_features parameter is null")
        if sparkContext is None:
            raise ValueError("sparkContext parameter is null")

        # prepare variables
        timestamp = "2019-01-01 00:00:00"
        output_dataset_id = str(
            configProperties.get("featureDatasetId"))
        tenant_id = str(
            configProperties.get("tenantId"))
        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 ['output_dataset_id', 'tenant_id', 'service_token', 'user_token', 'org_id', 'api_key']:
            if eval(arg) == 'None':
                raise ValueError("%s is empty" % arg)

        # create and prepare DataFrame with valid columns
        output_df = data_features.withColumn("date", col("date").cast(StringType()))
        output_df = output_df.withColumn(tenant_id, struct(col("date"), col("store"), col("features")))
        output_df = output_df.withColumn("timestamp", lit(timestamp).cast(TimestampType()))
        output_df = output_df.withColumn("_id", lit("empty"))
        output_df = output_df.withColumn("eventType", lit("empty"))

        # store data into dataset
        output_df.select(tenant_id, "_id", "eventType", "timestamp") \
            .write.format("com.adobe.platform.dataset") \
            .option('orgId', org_id) \
            .option('serviceToken', service_token) \
            .option('userToken', user_token) \
            .option('serviceApiKey', api_key) \
            .save(output_dataset_id)