End-to-end Machine Learning with Tensorflow on GCP

In this workshop, we walk through the process of building a complete machine learning pipeline covering ingest, exploration, training, evaluation, deployment, and prediction. Along the way, we will discuss how to explore and split large data sets correctly using BigQuery and notebooks. The machine learning model in TensorFlow will be developed on a small sample locally. The preprocessing operations will be implemented in Cloud Dataflow, so that the same preprocessing can be applied in streaming mode as well. The training of the model will then be distributed and scaled out on Cloud ML Engine. The trained model will be deployed as a microservice and predictions invoked from a web application.

This lab consists of 8 parts and will take you about 3 hours. It goes along with this slide deck.

What you need

To complete this lab, you need:

A project and bucket on Google Cloud Platform

What you learn

In this lab, you:

Explore a large dataset using Jupyter and BigQuery
Export data for machine learning using Cloud Dataflow
Develop a machine learning model in Tensorflow
Train a machine learning model at scale on Cloud ML Engine
Deploy the trained ML model as a microservice
Automate into a Kubeflow pipeline
Invoke the trained model from an AppEngine web application and a Java Dataflow pipeline.

This lab illustrates how you can carry out data exploration of large datasets, but continue to use familiar tools like Pandas and Jupyter. The "trick" is to do the first part of your aggregation in BigQuery, get back a Pandas dataset and then work with the smaller Pandas dataset locally. Google Cloud provides a managed Jupyter experience, so that you don't need to run notebook servers yourself.

Launch a notebook on GCP

To launch a notebook instance on GCP:

Step 1

Open Cloud Shell. The cloud shell icon is at the top right of the Google Cloud Platform web console:

Step 2

In Cloud Shell, type:

gcloud compute zones list

Pick a zone in a geographical region close to you.

Step 3

In Cloud Shell, type:

datalab create babyweight --zone <ZONE>

substituting the zone you chose in Step 2 for "<ZONE>". Datalab will take about 5 minutes to start. Move on to the next step while it continues working to setup your Jupyter server.

Invoke BigQuery

You will now use BigQuery, a serverless data warehouse, to explore the natality dataset so that we can choose the features for our machine learning model.

To invoke a BigQuery query:

Step 1

Navigate to the BigQuery console by selecting BigQuery from the top-left-corner ("hamburger") menu.

Step 2

In the BigQuery Console, click on Compose Query.

Step 3

In the query textbox, type:

#standardsql
SELECT
  plurality,
  COUNT(1) AS num_babies,
  AVG(weight_pounds) AS avg_wt
FROM
  publicdata.samples.natality
WHERE
  year > 2000 AND year < 2005
GROUP BY
  plurality

How many triplets were born in the US between 2000 and 2005? ___________

Draw graphs in Cloud Datalab

Switch back to your Cloud Shell window.

Step 1

If necessary, wait for Datalab to finish launching. Datalab is ready when you see a message prompting you to do a "Web Preview".

Step 2

Click on the Web Preview icon on the top-right corner of the Cloud Shell ribbon. Switch to port 8081.

Step 3

In Datalab, start a new notebook by clicking on the +Notebook icon.

Step 4

In a cell in Datalab, type the following, then click Run and wait until you see a table of data.

query="""
SELECT
  weight_pounds,
  is_male,
  mother_age,
  plurality,
  gestation_weeks
FROM
  publicdata.samples.natality
WHERE year > 2000
"""
import google.datalab.bigquery as bq
df = bq.Query(query + " LIMIT 100").execute().result().to_dataframe()
df.head()

Note that we have gotten the results from BigQuery as a Pandas dataframe.

Step 5

In the next cell in Datalab, type the following, then click Run

def get_distinct_values(column_name):
  sql = """
SELECT
  {0},
  COUNT(1) AS num_babies,
  AVG(weight_pounds) AS avg_wt
FROM
  publicdata.samples.natality
WHERE
  year > 2000
GROUP BY
  {0}
  """.format(column_name)
  return bq.Query(sql).execute().result().to_dataframe()

df = get_distinct_values('is_male')
df.plot(x='is_male', y='avg_wt', kind='bar');

Are male babies heavier or lighter than female babies? Did you know this? _______

Is the sex of the baby a good feature to use in our machine learning model? _____

Step 6

In the next cell in Datalab, type the following, then click Run

df = get_distinct_values('gestation_weeks')
df = df.sort_values('gestation_weeks')
df.plot(x='gestation_weeks', y='avg_wt', kind='bar');

This graph shows the average weight of babies born in the each week of pregancy. The way you'd read the graph is to look at the y-value for x=35 to find out the average weight of a baby born in the 35th week of pregnancy.

Is gestation_weeks a good feature to use in our machine learning model? _____

Is gestation_weeks always available? __________

Compare the variability of birth weight due to sex of baby and due to gestation weeks. Which factor do you think is more important for accurate weight prediction? __________________________________

Summary

In this step, you learned how to carry out data exploration of large datasets using BigQuery, Pandas, and Jupyter. The "trick" is to do the first part of your aggregation in BigQuery, get back a Pandas dataset and then work with the smaller Pandas dataset locally. Cloud Datalab provides a managed Jupyter experience, so that you don't need to run notebook servers yourself.

Clone repository

In Datalab:

Step 1

Click on the Datalab icon to go to the listing page and start a new notebook by clicking on the +Notebook icon.

Step 2

In the cell in Datalab, type the following, then click Run

%bash
git clone \
    https://github.com/GoogleCloudPlatform/training-data-analyst/

In this step, you learned to deploy a trained model as a microservice and get it to do both online and batch prediction.

Step 1

In CloudShell, type:

git clone \
    https://github.com/GoogleCloudPlatform/training-data-analyst/

cd training-data-analyst/training-data-analyst/tree/master/courses/machine_learning/deepdive/06_structured/pipelines

./create_cluster.sh

Step 2

Edit the following files in your favorite text editor so that the pipeline finishes quickly:

containers/hypertrain/train.sh	Change train_examples to 2000
containers/hypertrain/hyperparam.yaml	Change maxTrials to 2
Change maxParallelTrials to 2
containers/traintuned/	Change train_examples to 20000

Step 3

Navigate to the GKE section of the GCP console and make sure cluster is created.

Step 4

In CloudShell, build the containers:

cd containers
./build_all.sh
cd ..

Step 5

In CloudShell, start the UI:

./start_ui.sh

Step 6

Navigate to https://localhost:8085

Step 7

Open up a Terminal in the Jupyter environment

Step 8

Type in:

git clone \
    https://github.com/GoogleCloudPlatform/training-data-analyst/

Step 9

Go back to the listing page in Jupyter and navigate to: training-data-analyst/courses/machine_learning/deepdive/06_structured/pipelines and click on mlp_babyweight.py

Step 10

Edit the project name to match your project. It has to be a static string.

Step 11

Go back to the listing page in Jupyter and navigate to: training-data-analyst/courses/machine_learning/deepdive/06_structured/ and click on 7_pipelines.ipynb

Step 12

Run the notebook to upload the pipeline to Kubeflow pipelines.

Step 13

Go to the Kubeflow pipelines and create a run of the pipeline, changing project and bucket as necessary.

Step 14

Monitor the logs until done (will take about an hour).

Summary

In this step, you learned to deploy the entire training pipeline onto Kubeflow Pipelines and automate the training, hyperparameter tuning, and deployment of the model.

Step 1

Open CloudShell, and git clone the repository if necessary:

git clone \
    https://github.com/GoogleCloudPlatform/training-data-analyst/

Step 2

In CloudShell, deploy the website application:

cd training-data-analyst/courses/machine_learning/deepdive
cd 06_structured/serving
./deploy.sh

Step 3

In a browser, visit https://<PROJECT>.appspot.com/ and try out the application.

Step 4

In CloudShell, call a Java program that invokes the web service:

cd ~/training-data-analyst/courses/machine_learning/deepdive
cd 06_structured/serving
./run_once.sh

Step 5

In CloudShell, call a Dataflow pipeline that invokes the web service on a text file:

cd ~/training-data-analyst/courses/machine_learning/deepdive
cd 06_structured/serving
./run_ontext.sh

The code will also work real-time, reading from Pub/Sub and writing to BigQuery:

cd ~/training-data-analyst/courses/machine_learning/deepdive
cd 06_structured/serving
cat ./run_dataflow.sh

Summary

In this step, you deployed an AppEngine web application that consumes the machine learning service. You also looked at how to consume the ML predictions from Dataflow, both in batch mode and in real-time.

Step 1

Click on the person icon in the top-right corner of your Datalab window and click on the link to manage the VM.

Step 2

In the web console, select the Datalab VM and click DELETE.

Note: Cloud Datalab creates a a persistent disk to store your notebooks. These notebooks will be available the next time you start Datalab. Datalab also creates a firewall rule for tunneling. Use the web console to delete the disk and firewall rule if you will not use Datalab again.