What was made

What are you doing

Create a model
Deploy the model to GCP
Prediction on GCP

Titanic's "age" and "gender" are used as learning data.
We are making a two-class classification model to see if you can survive with XGBoost.
Deploy and predict the model on GCP's AI-Platform.

GoogleColab I made it a notebook, so if

Information and operation confirmation as of 2020/02/16.

https://colab.research.google.com/drive/1s6_o-nvMmHhAeBAOoRDgE0UNU6Fe3XrC

procedure

It has almost the same contents as notebook. If you are having trouble opening Colab, please see here.

GCP account registration

[Explanation with image] Register an account with a free trial of Google Cloud Platform (GCP)

Install Google Cloud SDK

Install Google Cloud SDK ~ Initialize

SDK certification

Authenticate with your Google account to mess with GCP using the gcloud command.

$ gcloud auth login

Project creation

ID is prohibited from fogging.

$ PROJECT_ID=anata-no-pj-id
$ PROJECT_NAME=anata-no-pj-name

$ gcloud projects create $PROJECT_ID \
--name $PROJECT_NAME

Billing account settings

If you do not set it, a 403 error will occur when accessing the bucket.

If the following pop-up does not appear, you can skip it because the billing account has already been set.

Activate the project

Set to the target project of command operation.

$ gcloud config set project $PROJECT_ID

Verification

! gcloud config list

# [component_manager]
# disable_update_check = True
# [compute]
# gce_metadata_read_timeout_sec = 0
# [core]
# account = [email protected]
# project = anata-no-pj-id
# 
# Your active configuration is: [default]

project = anata-no-project-id

If it is OK

Set region, zone, interpreter

REGION=us-central1
ZONE=us-central1-a

$ gcloud config set compute/region $REGION
$ gcloud config set compute/zone $ZONE
$ gcloud config set ml_engine/local_python $(which python3)

The regions where AI Platform online prediction can be used are as follows:

us-central1
us-east1
us-east4
asia-northeast1
europe-west1

The interpreter is specified to use python3 system for local training.

Verification

$ gcloud config list

# [component_manager]
# disable_update_check = True
# [compute]
# gce_metadata_read_timeout_sec = 0
# region = us-central1
# zone = us-central1-a
# [core]
# account = [email protected]
# project = anata-no-pj-id
# [ml_engine]
# local_python = /usr/bin/python3
# 
# Your active configuration is: [default]

region = us-central1
zone = us-central1-a
local_python = /usr/bin/python3

If it is OK

Clone a set of training code

https://github.com/komiyakomiyakomiya/titanic_prediction_on_gcp

$ git clone https://github.com/komiyakomiyakomiya/titanic_prediction_on_gcp.git

Create a directory to save the model

`notebook`


import os

os.makedirs('./titanic_prediction_on_gcp/working/models/', exist_ok=True)

Training & save model locally

The trained model is saved as ./titanic_prediction_on_gcp/working/models/model.pkl.

$ gcloud ai-platform local train \
--package-path titanic_prediction_on_gcp/working/ \
--module-name working.predict_xgb

Creating a bucket

Create a bucket in GCS to upload the saved model.

If you do not set a billing account, a 403 error will occur.

BUCKET_NAME=anata-no-bkt-name

$ gsutil mb -l $REGION gs://$BUCKET_NAME

Verification

$ gsutil ls -la

# gs://anata-no-bkt-name/

Upload the saved model to GCS

$ gsutil cp ./titanic_prediction_on_gcp/working/models/model.pkl gs://$BUCKET_NAME/models/model.pkl

Verification

$ gsutil ls gs://$BUCKET_NAME/models/

# gs://anata-no-bkt-name/models/model.pkl

Enable API

Enable the following two to use the AI-Platform API.

ml.googleapis.com
compute.googleapis.com

$ gcloud services enable ml.googleapis.com
$ gcloud services enable compute.googleapis.com

Verification

$ gcloud services list --enabled

# NAME                              TITLE
# bigquery.googleapis.com           BigQuery API
# bigquerystorage.googleapis.com    BigQuery Storage API
# cloudapis.googleapis.com          Google Cloud APIs
# clouddebugger.googleapis.com      Stackdriver Debugger API
# cloudtrace.googleapis.com         Stackdriver Trace API
# compute.googleapis.com            Compute Engine API
# datastore.googleapis.com          Cloud Datastore API
# logging.googleapis.com            Stackdriver Logging API
# ml.googleapis.com                 AI Platform Training & Prediction API
# monitoring.googleapis.com         Stackdriver Monitoring API
# oslogin.googleapis.com            Cloud OS Login API
# servicemanagement.googleapis.com  Service Management API
# serviceusage.googleapis.com       Service Usage API
# sql-component.googleapis.com      Cloud SQL
# storage-api.googleapis.com        Google Cloud Storage JSON API
# storage-component.googleapis.com  Cloud Storage

compute.googleapis.com Compute Engine API
ml.googleapis.com AI Platform Training & Prediction API

OK if there is

Creating model / version resources

Create a model resource and a version resource, and associate it with the uploaded model.pkl.

Model resource

MODEL_NAME=model_xgb
MODEL_VERSION=v1

$ gcloud ai-platform models create $MODEL_NAME \
--regions $REGION

Version resource

! gcloud ai-platform versions create $MODEL_VERSION \
--model $MODEL_NAME \
--origin gs://$BUCKET_NAME/models/ \
--runtime-version 1.14 \
--framework xgboost \
--python-version 3.5

Confirmation of input data

Let's make a prediction using the data prepared in advance. First, check the contents.

! cat titanic_prediction_on_gcp/input/titanic/predict.json

# [36.0, 0] <-36 years old,male

In the format [Age, Gender], the gender is male: 0, female: 1.

Predicted by AI-Platform

! gcloud ai-platform predict \
--model model_xgb \
--version $MODEL_VERSION \
--json-instances titanic_prediction_on_gcp/input/titanic/predict.json

[0.44441232085227966] It is OK if such a predicted value is returned.

Creating a service account

Next, access AI-Platform from python and get the prediction. You will need a service account key, so first create a service account.

SA_NAME=anata-no-sa-name
SA_DISPLAY_NAME=anata-no-sa-display-name

$ gcloud iam service-accounts create $SA_NAME \
--display-name $SA_DISPLAY_NAME \

Grant permissions to service accounts

$ gcloud projects add-iam-policy-binding $PROJECT_ID \
--member serviceAccount:[email protected] \
--role roles/iam.serviceAccountKeyAdmin

$ gcloud projects add-iam-policy-binding $PROJECT_ID \
--member serviceAccount:[email protected] \
--role roles/ml.admin

Service account key generation

$ gcloud iam service-accounts keys create titanic_prediction_on_gcp/service_account_keys/key.json \
--iam-account [email protected]

Read the path of key.json as an environment variable

Generate .env file and describe environment variables and paths

$ echo GOOGLE_APPLICATION_CREDENTIALS=/content/titanic_prediction_on_gcp/service_account_keys/key.json > /content/titanic_prediction_on_gcp/.env

Verification

$ cat ./titanic_prediction_on_gcp/.env

# GOOGLE_APPLICATION_CREDENTIALS=/content/titanic_prediction_on_gcp/service_account_keys/key.json

## Install python-dotenv I'm not in Colab

$ pip install python-dotenv

Define a function to get the prediction

`notebook`


import googleapiclient.discovery
from dotenv import load_dotenv

#Environment variable settings
load_dotenv('/content/titanic_prediction_on_gcp/.env')


def main(input_data):
    input_data = [input_data]

    PROJECT_ID = 'anata-no-pj-id'
    VERSION_NAME = 'v1'
    MODEL_NAME = 'model_xgb'

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

    response = service.projects().predict(
        name=name,
        body={'instances': input_data}
    ).execute()

    if 'error' in response:
        print(response['error'])
    else:
        pred = response['predictions'][0]

    return pred

Create a drop-down menu for age and gender

`notebook`


import ipywidgets as widgets
from ipywidgets import HBox, VBox


age = [i for i in range(101)]
sex = ['male', 'Female']

dropdown_age = widgets.Dropdown(options=age, description='age: ')
dropdown_sex = widgets.Dropdown(options=sex, description='sex: ')
variables = VBox(children=[dropdown_age, dropdown_sex])

VBox(children=[variables])

Forecast

`notebook`


import numpy as np
from IPython.display import Image
from IPython.display import display_png


input_age = float(dropdown_age.value)
input_sex = 0 if dropdown_sex.value == 'male' else 1
test_input = [input_age, input_sex]

pred = main(test_input)
# print(pred)
pred_binary = np.where(pred > 0.5, 1, 0)
# print(pred_binary)

print('\n When you ride a Titanic...')

if pred_binary == 1:
    display_png(Image('/content/titanic_prediction_on_gcp/images/alive.png'))
else:
    display_png(Image('/content/titanic_prediction_on_gcp/images/dead.png'))

Official reference

https://cloud.google.com/sdk/gcloud/reference/

https://cloud.google.com/sdk/gcloud/reference/ai-platform/

https://cloud.google.com/storage/docs/gsutil

Summary

Thank you for reading to the end. As a 36-year-old uncle, I'm about to die, so I'll be very careful when riding the Titanic.

If you ride the Titanic ... die.

What was made

What are you doing

procedure

GCP account registration

Install Google Cloud SDK

SDK certification

Project creation

Billing account settings

Activate the project

Verification

Set region, zone, interpreter

Verification

Clone a set of training code

Create a directory to save the model

notebook

Training & save model locally

Creating a bucket

Verification

Upload the saved model to GCS

Verification

Enable API

Verification

Creating model / version resources

Confirmation of input data

Predicted by AI-Platform

Creating a service account

Grant permissions to service accounts

Service account key generation

Read the path of key.json as an environment variable

Verification

Define a function to get the prediction

notebook

Create a drop-down menu for age and gender

notebook

Forecast

notebook

Official reference

Summary

`notebook`

`notebook`

`notebook`

`notebook`