Apache Beam RunInference for scikit-learn

Run in Google Colab View source on GitHub

This notebook demonstrates the use of the RunInference transform for scikit-learn, also called sklearn. Apache Beam RunInference has implementations of the ModelHandler class prebuilt for scikit-learn. For more information about using RunInference, see Get started with AI/ML pipelines in the Apache Beam documentation.

You can choose the appropriate model handler based on your input data type:

With RunInference, these model handlers manage batching, vectorization, and prediction optimization for your scikit-learn pipeline or model.

This notebook demonstrates the following common RunInference patterns:

  • Generate predictions.
  • Postprocess results after RunInference.
  • Run inference with multiple models in the same pipeline.

The linear regression models used in these samples are trained on data that correspondes to the 5 and 10 times tables; that is,y = 5x and y = 10x respectively.

Before you begin

Complete the following setup steps:

  1. Install dependencies for Apache Beam.
  2. Authenticate with Google Cloud.
  3. Specify your project and bucket. You use the project and bucket to save and load models.
pip install google-api-core --quiet
pip install google-cloud-pubsub google-cloud-bigquery-storage --quiet
pip install apache-beam[gcp,dataframe] --quiet

About scikit-learn versions

scikit-learn is a build-dependency of Apache Beam. If you need to install a different version of sklearn , use %pip install scikit-learn==<version>

from google.colab import auth
auth.authenticate_user()

import pickle
from sklearn import linear_model
from typing import Tuple

import numpy as np
import apache_beam as beam

from apache_beam.ml.inference.sklearn_inference import ModelFileType
from apache_beam.ml.inference.sklearn_inference import SklearnModelHandlerNumpy
from apache_beam.ml.inference.base import KeyedModelHandler
from apache_beam.ml.inference.base import PredictionResult
from apache_beam.ml.inference.base import RunInference
from apache_beam.options.pipeline_options import PipelineOptions

# NOTE: If an error occurs, restart your runtime.

import os

# Constants
project = "<PROJECT_ID>"
bucket = "<BUCKET_NAME>" 

# To avoid warnings, set the project.
os.environ['GOOGLE_CLOUD_PROJECT'] = project

Create the data and the scikit-learn model

This section demonstrates the following steps:

  1. Create the data to train the scikit-learn linear regression model.
  2. Train the linear regression model.
  3. Save the scikit-learn model using pickle.

In this example, you create two models, one with the 5 times model and a second with the 10 times model.

# Input data to train the sklearn model for the 5 times table.
x = np.arange(0, 100, dtype=np.float32).reshape(-1, 1)
y = (x * 5).reshape(-1, 1)

def train_and_save_model(x, y, model_file_name):
  regression = linear_model.LinearRegression()
  regression.fit(x,y)

  with open(model_file_name, 'wb') as f:
      pickle.dump(regression, f)

five_times_model_filename = 'sklearn_5x_model.pkl'
train_and_save_model(x, y, five_times_model_filename)

# Change y to be 10 times, and output a 10 times table.
ten_times_model_filename = 'sklearn_10x_model.pkl'
train_and_save_model(x, y, ten_times_model_filename)
y = (x * 10).reshape(-1, 1)
train_and_save_model(x, y, 'sklearn_10x_model.pkl')

Create a scikit-learn RunInference pipeline

This section demonstrates how to do the following:

  1. Define a scikit-learn model handler that accepts an array_like object as input.
  2. Read the data from BigQuery.
  3. Use the scikit-learn trained model and the scikit-learn RunInference transform on unkeyed data.
%pip install --upgrade google-cloud-bigquery --quiet

gcloud config set project $project
 
Updated property [core/project].

# Populated BigQuery table

from google.cloud import bigquery

client = bigquery.Client(project=project)

# Make sure the dataset_id is unique in your project.
dataset_id = '{project}.maths'.format(project=project)
dataset = bigquery.Dataset(dataset_id)

# Modify the location based on your project configuration.
dataset.location = 'US'
dataset = client.create_dataset(dataset, exists_ok=True)

# Table name in the BigQuery dataset.
table_name = 'maths_problems_1'

query = """
    CREATE OR REPLACE TABLE
      {project}.maths.{table} ( key STRING OPTIONS(description="A unique key for the maths problem"),
    value FLOAT64 OPTIONS(description="Our maths problem" ) );
    INSERT INTO maths.{table}
    VALUES
      ("first_example", 105.00),
      ("second_example", 108.00),
      ("third_example", 1000.00),
      ("fourth_example", 1013.00)
""".format(project=project, table=table_name)

create_job = client.query(query)
create_job.result()

<google.cloud.bigquery.table._EmptyRowIterator at 0x7f97abb4e850>

sklearn_model_handler = SklearnModelHandlerNumpy(model_uri=five_times_model_filename) 


pipeline_options = PipelineOptions().from_dictionary(
                                      {'temp_location':f'gs://{bucket}/tmp'})

# Define the BigQuery table specification.
table_name = 'maths_problems_1'
table_spec = f'{project}:maths.{table_name}'

with beam.Pipeline(options=pipeline_options) as p:
  (
      p 
      | "ReadFromBQ" >> beam.io.ReadFromBigQuery(table=table_spec)
      | "ExtractInputs" >> beam.Map(lambda x: [x['value']]) 
      | "RunInferenceSklearn" >> RunInference(model_handler=sklearn_model_handler)
      | beam.Map(print)
  )

PredictionResult(example=[1000.0], inference=array([5000.]))
PredictionResult(example=[1013.0], inference=array([5065.]))
PredictionResult(example=[108.0], inference=array([540.]))
PredictionResult(example=[105.0], inference=array([525.]))

Use sklearn RunInference on keyed inputs

This section demonstrates how to do the following:

  1. Wrap the SklearnModelHandlerNumpy object around KeyedModelHandler to handle keyed data.
  2. Read the data from BigQuery.
  3. Use the sklearn trained model and the sklearn RunInference transform on a keyed data.
sklearn_model_handler = SklearnModelHandlerNumpy(model_uri=five_times_model_filename) 
keyed_sklearn_model_handler = KeyedModelHandler(sklearn_model_handler)

pipeline_options = PipelineOptions().from_dictionary(
                                      {'temp_location':f'gs://{bucket}/tmp'})
with beam.Pipeline(options=pipeline_options) as p:
  (
  p 
  | "ReadFromBQ" >> beam.io.ReadFromBigQuery(table=table_spec)
  | "ExtractInputs" >> beam.Map(lambda x: (x['key'], [x['value']])) 
  | "RunInferenceSklearn" >> RunInference(model_handler=keyed_sklearn_model_handler)
  | beam.Map(print)
  )

('third_example', PredictionResult(example=[1000.0], inference=array([5000.])))
('fourth_example', PredictionResult(example=[1013.0], inference=array([5065.])))
('second_example', PredictionResult(example=[108.0], inference=array([540.])))
('first_example', PredictionResult(example=[105.0], inference=array([525.])))

Run multiple models

This code creates a pipeline that takes two RunInference transforms with different models and then combines the output.

from typing import Tuple

def format_output(run_inference_output) -> str:
  """Takes input from RunInference for scikit-learn and extracts the output."""
  key, prediction_result = run_inference_output
  example = prediction_result.example[0]
  prediction = prediction_result.inference[0]
  return f"key = {key}, example = {example} -> predictions {prediction}"

five_times_model_handler = KeyedModelHandler(
    SklearnModelHandlerNumpy(model_uri=five_times_model_filename))
ten_times_model_handler = KeyedModelHandler(
    SklearnModelHandlerNumpy(model_uri=ten_times_model_filename))

pipeline_options = PipelineOptions().from_dictionary(
                                      {'temp_location':f'gs://{bucket}/tmp'})
with beam.Pipeline(options=pipeline_options) as p:
  inputs = (p 
    | "ReadFromBQ" >> beam.io.ReadFromBigQuery(table=table_spec))
  five_times = (inputs
    | "Extract For 5" >> beam.Map(lambda x: ('{} {}'.format(x['key'], '* 5'), [x['value']]))
    | "5 times" >> RunInference(model_handler = five_times_model_handler))
  ten_times = (inputs
    | "Extract For 10" >> beam.Map(lambda x: ('{} {}'.format(x['key'], '* 10'), [x['value']]))
    | "10 times" >> RunInference(model_handler = ten_times_model_handler))
  _ = ((five_times, ten_times) | "Flattened" >> beam.Flatten()
    | "format output" >> beam.Map(format_output)
    | "Print" >> beam.Map(print))

key = third_example * 10, example = 1000.0 -> predictions 10000.0
key = fourth_example * 10, example = 1013.0 -> predictions 10130.0
key = second_example * 10, example = 108.0 -> predictions 1080.0
key = first_example * 10, example = 105.0 -> predictions 1050.0
key = third_example * 5, example = 1000.0 -> predictions 5000.0
key = fourth_example * 5, example = 1013.0 -> predictions 5065.0
key = second_example * 5, example = 108.0 -> predictions 540.0
key = first_example * 5, example = 105.0 -> predictions 525.0