StatsForecast ETS 和 Facebook Prophet 在 Spark 上的应用 (M5)

pip install prophet "neuralforecast<1.0.0" "statsforecast[fugue]"

from time import time

from neuralforecast.data.datasets.m5 import M5, M5Evaluation
from statsforecast.distributed.utils import forecast
from statsforecast.distributed.fugue import FugueBackend
from statsforecast.models import ETS, SeasonalNaive
from statsforecast.core import StatsForecast

from pyspark.sql import SparkSession

spark = SparkSession.builder.getOrCreate()
backend = FugueBackend(spark, {"fugue.spark.use_pandas_udf":True})

init = time()
ets_forecasts = backend.forecast(
    "s3://m5-benchmarks/data/train/m5-target.parquet", 
    [ETS(season_length=7, model='ZAA')], 
    freq="D", 
    h=28, 
).toPandas()
end = time()
print(f'Minutes taken by StatsForecast on a Spark cluster: {(end - init) / 60}')

Y_hat = ets_forecasts.set_index(['unique_id', 'ds']).unstack()
Y_hat = Y_hat.droplevel(0, 1).reset_index()

*_, S_df = M5.load('./data')
Y_hat = S_df.merge(Y_hat, how='left', on=['unique_id'])#.drop(columns=['unique_id'])

wrmsse_ets = M5Evaluation.evaluate(y_hat=Y_hat, directory='./data')

wrmsse_ets

import logging
from time import time

import pandas as pd
from neuralforecast.data.datasets.m5 import M5, M5Evaluation
from prophet import Prophet
from pyspark.sql.types import *

# disable informational messages from prophet
logging.getLogger('py4j').setLevel(logging.ERROR)

# structure of the training data set
train_schema = StructType([
  StructField('unique_id', StringType()),  
  StructField('ds', DateType()),
  StructField('y', DoubleType())
  ])
 
# read the training file into a dataframe
train = spark.read.parquet(
  's3://m5-benchmarks/data/train/m5-target.parquet', 
  header=True, 
  schema=train_schema
 )
 
# make the dataframe queriable as a temporary view
train.createOrReplaceTempView('train')

sql_statement = '''
  SELECT
    unique_id AS unique_id,
    CAST(ds as date) as ds,
    y as y
  FROM train
  '''
 
m5_history = (
  spark
    .sql( sql_statement )
    .repartition(sc.defaultParallelism, ['unique_id'])
  ).cache()

def forecast( history_pd: pd.DataFrame ) -> pd.DataFrame:
  
  # TRAIN MODEL AS BEFORE
  # --------------------------------------
  # remove missing values (more likely at day-store-item level)
    history_pd = history_pd.dropna()

    # configure the model
    model = Prophet(
        growth='linear',
        daily_seasonality=False,
        weekly_seasonality=True,
        yearly_seasonality=True,
        seasonality_mode='multiplicative'
    )

    # train the model
    model.fit( history_pd )
    # --------------------------------------

    # BUILD FORECAST AS BEFORE
    # --------------------------------------
    # make predictions
    future_pd = model.make_future_dataframe(
        periods=28, 
        freq='d', 
        include_history=False
    )
    forecast_pd = model.predict( future_pd )  
    # --------------------------------------

    # ASSEMBLE EXPECTED RESULT SET
    # --------------------------------------
    # get relevant fields from forecast
    forecast_pd['unique_id'] = history_pd['unique_id'].unique()[0]
    f_pd = forecast_pd[['unique_id', 'ds','yhat']]
    # --------------------------------------

    # return expected dataset
    return f_pd

result_schema = StructType([
  StructField('unique_id', StringType()), 
  StructField('ds',DateType()),
  StructField('yhat',FloatType()),
])

init = time()
results = (
  m5_history
    .groupBy('unique_id')
      .applyInPandas(forecast, schema=result_schema)
    ).toPandas()
end = time()
print(f'Minutes taken by Prophet on a Spark cluster: {(end - init) / 60}')

Y_hat = results.set_index(['unique_id', 'ds']).unstack()
Y_hat = Y_hat.droplevel(0, 1).reset_index()

*_, S_df = M5.load('./data')
Y_hat = S_df.merge(Y_hat, how='left', on=['unique_id'])#.drop(columns=['unique_id'])

wrmsse = M5Evaluation.evaluate(y_hat=Y_hat, directory='./data')

wrmsse