In [ ]:
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# See the License for the specific language governing permissions and
# limitations under the License.

Online Prediction with scikit-learn on AI Platform

This notebook uses the Census Income Data Set to create a simple model, train the model, upload the model to Ai Platform, and lastly use the model to make predictions.

How to bring your model to AI Platform

Getting your model ready for predictions can be done in 5 steps:

  1. Save your model to a file
  2. Upload the saved model to Google Cloud Storage
  3. Create a model resource on AI Platform
  4. Create a model version (linking your scikit-learn model)
  5. Make an online prediction


Before you jump in, let’s cover some of the different tools you’ll be using to get online prediction up and running on AI Platform.

Google Cloud Platform lets you build and host applications and websites, store data, and analyze data on Google's scalable infrastructure.

AI Platform is a managed service that enables you to easily build machine learning models that work on any type of data, of any size.

Google Cloud Storage (GCS) is a unified object storage for developers and enterprises, from live data serving to data analytics/ML to data archiving.

Cloud SDK is a command line tool which allows you to interact with Google Cloud products. In order to run this notebook, make sure that Cloud SDK is installed in the same environment as your Jupyter kernel.

Part 0: Setup

These variables will be needed for the following steps.


  • PROJECT_ID <YOUR_PROJECT_ID> - with your project's id. Use the PROJECT_ID that matches your Google Cloud Platform project.
  • BUCKET_NAME <YOUR_BUCKET_NAME> - with the bucket id you created above.
  • MODEL_NAME <YOUR_MODEL_NAME> - with your model name, such as 'census'
  • VERSION <YOUR_VERSION> - with your version name, such as 'v1'
  • REGION <REGION> - select a region or use the default 'us-central1'. The region is where the model will be deployed.

In [1]:
%env MODEL_NAME census
%env REGION us-central1

env: MODEL_NAME=census
env: REGION=us-central1

Download the data

The Census Income Data Set that this sample uses for training is hosted by the UC Irvine Machine Learning Repository.

  • Training file is
  • Evaluation file is adult.test


This dataset is provided by a third party. Google provides no representation, warranty, or other guarantees about the validity or any other aspects of this dataset.

In [2]:
# Create a directory to hold the data
! mkdir census_data

# Download the data
! curl --output census_data/
! curl --output census_data/adult.test

mkdir: census_data: File exists
  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100 3881k  100 3881k    0     0  7804k      0 --:--:-- --:--:-- --:--:-- 7824k
  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100 1956k  100 1956k    0     0  6895k      0 --:--:-- --:--:-- --:--:-- 6912k

Part 1: Train/Save the model

First, the data is loaded into a pandas DataFrame that can be used by scikit-learn. Then a simple model is created and fit against the training data. Lastly, sklearn's built in version of joblib is used to save the model to a file that can be uploaded to AI Platform.

In [3]:
import googleapiclient.discovery
import json
import numpy as np
import os
import pandas as pd
import pickle

from sklearn.ensemble import RandomForestClassifier
from sklearn.externals import joblib
from sklearn.feature_selection import SelectKBest
from sklearn.pipeline import FeatureUnion
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import LabelBinarizer

# Define the format of your input data including unused columns (These are the columns from the census data files)

# Categorical columns are columns that need to be turned into a numerical value to be used by scikit-learn

# Load the training census dataset
with open('./census_data/', 'r') as train_data:
    raw_training_data = pd.read_csv(train_data, header=None, names=COLUMNS)

# Remove the column we are trying to predict ('income-level') from our features list
# Convert the Dataframe to a lists of lists
train_features = raw_training_data.drop('income-level', axis=1).as_matrix().tolist()
# Create our training labels list, convert the Dataframe to a lists of lists
train_labels = (raw_training_data['income-level'] == ' >50K').as_matrix().tolist()

# Load the test census dataset
with open('./census_data/adult.test', 'r') as test_data:
    raw_testing_data = pd.read_csv(test_data, names=COLUMNS, skiprows=1)
# Remove the column we are trying to predict ('income-level') from our features list
# Convert the Dataframe to a lists of lists
test_features = raw_testing_data.drop('income-level', axis=1).values.tolist()
# Create our training labels list, convert the Dataframe to a lists of lists
test_labels = (raw_testing_data['income-level'] == ' >50K.').values.tolist()

# Since the census data set has categorical features, we need to convert
# them to numerical values. We'll use a list of pipelines to convert each
# categorical column and then use FeatureUnion to combine them before calling
# the RandomForestClassifier.
categorical_pipelines = []

# Each categorical column needs to be extracted individually and converted to a numerical value.
# To do this, each categorical column will use a pipeline that extracts one feature column via
# SelectKBest(k=1) and a LabelBinarizer() to convert the categorical value to a numerical one.
# A scores array (created below) will select and extract the feature column. The scores array is
# created by iterating over the COLUMNS and checking if it is a CATEGORICAL_COLUMN.
for i, col in enumerate(COLUMNS[:-1]):
        # Create a scores array to get the individual categorical column.
        # Example:
        #  data = [39, 'State-gov', 77516, 'Bachelors', 13, 'Never-married', 'Adm-clerical', 
        #         'Not-in-family', 'White', 'Male', 2174, 0, 40, 'United-States']
        #  scores = [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
        # Returns: [['State-gov']]
        # Build the scores array.
        scores = [0] * len(COLUMNS[:-1]) 
        # This column is the categorical column we want to extract.
        scores[i] = 1
        skb = SelectKBest(k=1)
        skb.scores_ = scores
        # Convert the categorical column to a numerical value
        lbn = LabelBinarizer()
        r = skb.transform(train_features)
        # Create the pipeline to extract the categorical feature
            ('categorical-{}'.format(i), Pipeline([
                ('SKB-{}'.format(i), skb),
                ('LBN-{}'.format(i), lbn)])))

# Create pipeline to extract the numerical features
skb = SelectKBest(k=6)
# From COLUMNS use the features that are numerical
skb.scores_ = [1, 0, 1, 0, 1, 0, 0, 0, 0, 0, 1, 1, 1, 0]
categorical_pipelines.append(('numerical', skb))

# Combine all the features using FeatureUnion
preprocess = FeatureUnion(categorical_pipelines)

# Create the classifier
classifier = RandomForestClassifier()

# Transform the features and fit them to the classifier, train_labels)

# Create the overall model as a single pipeline
pipeline = Pipeline([
    ('union', preprocess),
    ('classifier', classifier)

# Export the model to a file
joblib.dump(pipeline, 'model.joblib')

print('Model trained and saved')

Model trained and saved

Part 2: Upload the model

Next, you'll need to upload the model to your project's storage bucket in GCS. To use your model with AI Platform, it needs to be uploaded to Google Cloud Storage (GCS). This step takes your local ‘model.joblib’ file and uploads it GCS via the Cloud SDK using gsutil.

Before continuing, make sure you're properly authenticated and have access to the bucket. This next command sets your project to the one specified above.

Note: If you get an error below, make sure the Cloud SDK is installed in the kernel's environment.

In [4]:
! gcloud config set project $PROJECT_ID

Updated property [core/project].

Note: The exact file name of of the exported model you upload to GCS is important! Your model must be named “model.joblib”, “model.pkl”, or “model.bst” with respect to the library you used to export it. This restriction ensures that the model will be safely reconstructed later by using the same technique for import as was used during export.

In [5]:
! gsutil cp ./model.joblib gs://$BUCKET_NAME/model.joblib

Copying file://./model.joblib [Content-Type=application/octet-stream]...
| [1 files][  7.8 MiB/  7.8 MiB]                                                
Operation completed over 1 objects/7.8 MiB.                                      

Part 3: Create a model resource

AI Platform organizes your trained models using model and version resources. An AI Platform model is a container for the versions of your machine learning model. For more information on model resources and model versions look here.

At this step, you create a container that you can use to hold several different versions of your actual model.

In [6]:
! gcloud ml-engine models create $MODEL_NAME --regions $REGION

Created AI Platform model [projects/PROJECT_ID/models/census].

Part 4: Create a model version

Now it’s time to get your model online and ready for predictions. The model version requires a few components as specified here.

  • name - The name specified for the version when it was created. This will be the VERSION_NAME variable you declared at the beginning.
  • model - The name of the model container we created in Part 3. This is the MODEL_NAME variable you declared at the beginning.
  • deployment Uri - The Google Cloud Storage location of the trained model used to create the version. This is the bucket that you uploaded the model to with your BUCKET_NAME
  • runtime version - Select Google Cloud ML runtime version to use for this deployment. This is set to 1.4
  • framework - The framework specifies if you are using: TENSORFLOW, SCIKIT_LEARN, XGBOOST. This is set to SCIKIT_LEARN
  • pythonVersion - This specifies whether you’re using Python 2.7 or Python 3.5. The default value is set to “2.7”, if you are using Python 3.5, set the value to “3.5”

Note: If you require a feature of scikit-learn that isn’t available in the publicly released version yet, you can specify “runtimeVersion”: “HEAD” instead, and that would get the latest version of scikit-learn available from the github repo. Otherwise the following versions will be used:

  • scikit-learn: 0.19.0

First, we need to create a YAML file to configure our model version.


In [7]:
%%writefile ./config.yaml
deploymentUri: "gs://BUCKET_NAME/"
runtimeVersion: '1.4'
framework: "SCIKIT_LEARN"
pythonVersion: "3.5"

Writing ./config.yaml

Use the created YAML file to create a model version.

Note: It can take several minutes for you model to be available.

In [8]:
! gcloud ml-engine versions create $VERSION_NAME \
    --model $MODEL_NAME \
    --config config.yaml

Creating version (this might take a few minutes)......done.                    

Part 5: Make an online prediction

It’s time to make an online prediction with your newly deployed model. Before you begin, you'll need to take some of the test data and prepare it, so that the test data can be used by the deployed model.

In [9]:
# Get one person that makes <=50K and one that makes >50K to test our model.
print('Show a person that makes <=50K:')
print('\tFeatures: {0} --> Label: {1}\n'.format(test_features[0], test_labels[0]))

with open('less_than_50K.json', 'w') as outfile:
  json.dump(test_features[0], outfile)

print('Show a person that makes >50K:')
print('\tFeatures: {0} --> Label: {1}'.format(test_features[3], test_labels[3]))

with open('more_than_50K.json', 'w') as outfile:
  json.dump(test_features[3], outfile)

Show a person that makes <=50K:
	Features: [25, ' Private', 226802, ' 11th', 7, ' Never-married', ' Machine-op-inspct', ' Own-child', ' Black', ' Male', 0, 0, 40, ' United-States'] --> Label: False

Show a person that makes >50K:
	Features: [44, ' Private', 160323, ' Some-college', 10, ' Married-civ-spouse', ' Machine-op-inspct', ' Husband', ' Black', ' Male', 7688, 0, 40, ' United-States'] --> Label: True

Use gcloud to make online predictions

Use the two people (as seen in the table) gathered in the previous step for the gcloud predictions.

Person age workclass fnlwgt education education-num marital-status occupation
1 25 Private 226802 11th 7 Never-married Machine-op-inspect
2 44 Private 160323 Some-college 10 Married-civ-spouse Machine-op-inspct
Person relationship race sex capital-gain capital-loss hours-per-week native-country (Label) income-level
1 Own-child Black Male 0 0 40 United-States False (<=50K)
2 Huasband Black Male 7688 0 40 United-States True (>50K)

Test the model with an online prediction using the data of a person who makes <=50K.

Note: If you see an error, the model from Part 4 may not be created yet as it takes several minutes for a new model version to be created.

In [10]:
! gcloud ml-engine predict --model $MODEL_NAME --version $VERSION_NAME --json-instances less_than_50K.json


Test the model with an online prediction using the data of a person who makes >50K.

In [11]:
! gcloud ml-engine predict --model $MODEL_NAME --version $VERSION_NAME --json-instances more_than_50K.json


Use Python to make online predictions

Test the model with the entire test set and print out some of the results.

Note: If running notebook server on Compute Engine, make sure to "allow full access to all Cloud APIs".

In [12]:
import googleapiclient.discovery
import os
import pandas as pd

PROJECT_ID = os.environ['PROJECT_ID']
MODEL_NAME = os.environ['MODEL_NAME']

service ='ml', 'v1')
name = 'projects/{}/models/{}'.format(PROJECT_ID, MODEL_NAME)
name += '/versions/{}'.format(VERSION_NAME)

# Due to the size of the data, it needs to be split in 2
first_half = test_features[:int(len(test_features)/2)]
second_half = test_features[int(len(test_features)/2):]

complete_results = []
for data in [first_half, second_half]:
    responses = service.projects().predict(
        body={'instances': data}

    if 'error' in responses:
# Print the first 10 responses
for i, response in enumerate(complete_results[:10]):
    print('Prediction: {}\tLabel: {}'.format(response, test_labels[i]))

Prediction: False	Label: False
Prediction: False	Label: False
Prediction: False	Label: True
Prediction: True	Label: True
Prediction: False	Label: False
Prediction: False	Label: False
Prediction: False	Label: False
Prediction: False	Label: True
Prediction: False	Label: False
Prediction: False	Label: False

[Optional] Part 6: Verify Results

Use a confusion matrix to create a visualization of the online predicted results from AI Platform.

In [13]:
actual = pd.Series(test_labels, name='actual')
online = pd.Series(complete_results, name='online')


online False True
False 11606 829
True 1644 2202

Use a confusion matrix create a visualization of the predicted results from the local model. These results should be identical to the results above.

In [14]:
local_results = pipeline.predict(test_features)
local = pd.Series(local_results, name='local')


local False True
False 11606 829
True 1644 2202

Directly compare the two results

In [15]:
identical = 0
different = 0

for i in range(len(complete_results)):
    if complete_results[i] == local_results[i]:
        identical += 1
        different += 1
print('identical: {}, different: {}'.format(identical,different))

identical: 16281, different: 0

If all results are identical, it means you've successfully uploaded your local model to AI Platform and performed online predictions correctly.