Where ml ai_heavy

© 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved. @jrhunt
Randall Hunt – Some Guy From Los Angeles
@WhereML a Serverless AI Powered Location
Guessing Twitter Bot
Built with Amazon SageMaker and AWS Lambda
Based on LocationNet work by Jaeyoung Choi and Kevin Li

About Me
• Technical Evangelist at AWS
• I build some demos: https://github.com/ranman
• I write some blogs: https://aws.amazon.com/blogs/aws/author/randhunt/
• Formerly of SpaceX, NASA, MongoDB
• I like Python
• I dislike javascript
• I look ridiculous in my badge photo

Example
Try it! Tweet to
@WhereML with a picture.
Hold your cell phone
camera to the screen

Architecture
AWS Lambda FunctionAmazon API
Gateway
Amazon SageMaker
Model Artifacts
Inference
Endpoint
Inference code
Amazon ECR
Inference code

© 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Solving Some Of The Hardest Problems In Computer Science
Learning Language Perception Problem
Solving
Reasoning

The (60 year) rise of Artificial Intelligence

Put machine learning in the hands of every developer
and data scientist
ML @ AWS: Our mission

Customer Running ML on AWS Today

© 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved. @jrhunt© 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Reviewing The ML Process

Data Visualization &
Analysis
Business Problem –
ML problem framing Data Collection
Data Integration
Data Preparation &
Cleaning
Feature Engineering
Model Training &
Parameter Tuning
Model Evaluation
Are Business
Goals met?
Model Deployment
Monitoring &
Debugging
– Predictions
YesNo
DataAugmentation
Feature
Augmentation
The Machine Learning Process
Re-training

Analysis
Data Integration
Data Preparation &
Cleaning
Feature Engineering
Model Training &
Parameter Tuning
Model Evaluation
Are Business
Goals met?
Model Deployment
Monitoring &
Debugging
– Predictions
YesNo
DataAugmentation
Feature
Augmentation
Discovery: The Analysts
Re-training
• Help formulate the right
questions
• Domain Knowledge

Analysis
Data Integration
Data Preparation &
Cleaning
Feature Engineering
Model Training &
Parameter Tuning
Model Evaluation
Are Business
Goals met?
Model Deployment
Monitoring &
Debugging
– Predictions
YesNo
DataAugmentation
Feature
Augmentation
Integration: The Data Architecture
Retraining
• Build the data platform:
• Amazon S3
• AWS Glue
• Amazon Athena
• Amazon EMR
• Amazon Redshift
Spectrum

Analysis
Feature Engineering
Model Training &
Parameter Tuning
Model Evaluation
• Setup and manage
Notebook Environments
• Setup and manage
Training Clusters
• Write Data Connectors
• Scale ML algorithms to
large datasets
• Distribute ML training
algorithm to multiple
machines
• Secure Model artifacts
Why We built Amazon SageMaker: The Model Training Undifferentiated Heavy Lifting

Model Deployment
Monitoring &
Debugging
– Predictions
• Setup and manage Model
Inference Clusters
• Manage and Scale Model
Inference APIs
• Monitor and Debug Model
Predictions
• Models versioning and
performance tracking
• Automate New Model
version promotion to
production (A/B testing)
Why We built Amazon SageMaker: The Model Deployment Undifferentiated Heavy Lifting

A fully managed service that enables data scientists and
developers to quickly and easily build machine-learning
based models into production smart applications.
Amazon SageMaker

Amazon SageMaker
1 2 3 4
I I I I
Notebook Instances Algorithms ML Training Service ML Hosting Service

1
I
Notebook Instances
Zero Setup For Exploratory Data Analysis
Authoring &
Notebooks
ETL Access to AWS
Database services
Access to S3 Data
Lake
• Recommendations/Personalization
• Fraud Detection
• Forecasting
• Image Classification
• Churn Prediction
• Marketing Email/Campaign Targeting
• Log processing and anomaly detection
• Speech to Text
• More…
“Just add data”

Streaming datasets,
for cheaper training
Train faster, in a
single pass
Greater reliability on
extremely large
datasets
Choice of several ML
algorithms
Amazon SageMaker: 10x better algorithms
2
I
Algorithms

Cost vs. Time
$$$$
$$$
$$
$
Minutes Hours Days Weeks Months
Single
Machine
Distributed, with
Strong Machines

Infinitely Scalable ML Algorithms

2
I
Algorithms
Training code
• Matrix Factorization
• Regression
• Principal Component Analysis
• K-Means Clustering
• Gradient Boosted Trees
• And More!
Amazon provided Algorithms
Bring Your Own Script (IM builds the Container)
IM Estimators in
Apache Spark Bring Your Own Algorithm (You build the Container)
Amazon SageMaker: 10x better algorithms

Managed Distributed Training with Flexibility
Training code
• Matrix Factorization
• Regression
• Principal Component Analysis
• K-Means Clustering
• Gradient Boosted Trees
• And More!
Amazon provided Algorithms
Bring Your Own Script (IM builds the Container)
Bring Your Own Algorithm (You build the Container)
3
I
ML Training Service
Fetch Training data
Save Model Artifacts
Fully
managed –
Secured–
Amazon ECR
Save Inference Image
IM Estimators in
Apache Spark
CPU GPU HPO

4
I
ML Hosting Service
Amazon ECR
Amazon SageMaker
Easy Model Deployment to Amazon SageMaker
Versions of the same
inference code saved in
inference containers.
Prod is the primary one,
50% of the traffic must
be served there!

4
I
ML Hosting Service
Amazon ECR
Model Artifacts
Inference Image
be served there!
Create a Model
ModelName: prod
Amazon SageMaker

4
I
ML Hosting Service
Amazon ECR
Model Artifacts
Inference Image
Model versions
be served there!
Create versions of a Model
Amazon SageMaker

4
I
ML Hosting Service
Amazon ECR
30 50
10 10
InstanceType: c3.4xlarge
InitialInstanceCount: 3
ModelName: prod
VariantName: primary
InitialVariantWeight: 50
ProductionVariant
Model Artifacts
Inference Image
Model versions
be served there!
Create weighted
ProductionVariants
Amazon SageMaker

4
I
ML Hosting Service
Amazon ECR
30 50
10 10
ProductionVariant
Model Artifacts
Inference Image
Model versions
be served there!
Create an
EndpointConfiguration from
one or many
ProductionVariant(s)EndpointConfiguration
Amazon SageMaker
ModelName: prod

4
I
ML Hosting Service
Amazon ECR
30 50
10 10
ProductionVariant
Model Artifacts
Inference Image
Model versions
be served there! Create an Endpoint from
one EndpointConfiguration
EndpointConfiguration
Inference Endpoint
Amazon SageMaker
ModelName: prod

4
I
ML Hosting Service
Amazon ECR
30 50
10 10
ProductionVariant
Model Artifacts
Inference Image
Model versions
be served there!
One-Click!
EndpointConfiguration
Inference Endpoint
Amazon Provided Algorithms
Amazon SageMaker
ModelName: prod

4
I
ML Hosting Service
 Auto-Scaling Inference
APIs
 A/B Testing (more to
come)
 Low Latency & High
Throughput
 Bring Your Own Model
 Python SDK
Amazon SageMaker

Building the Model

Credits
• This model, LocationNet, was built by Jaeyoung Choi of
the International Computer Science Institute and Kevin
Li of the University of California, Berkley
• Supported by the AWS Cloud Credits for Research
Program
• Based on work on PlaNet by Weyland et. all
Ref: arxiv.org/abs/1602.05314 : PlaNet—Photo
Geolocation with Convolutional Neural Networks

LocationNet Model Approach
• Model trained and built with Apache MXNet
• Trained with 33.9 million geo-tagged images from the
AWS Multimedia Commons Dataset for 12 epochs over
9 days using a single p2.16xlarge.
• Uses Google’s S2 Spherical Geometry library to
subdivide the earth into 15,527 of multi-scale geographic
cells which serve as classes for the data.
• Built on ResNet-101 architecture

Model Architecture

Example S2 Multi-scale partitioning
Ref: arxiv.org/abs/1602.05314 : PlaNet—Photo Geolocation with
Convolutional Neural Networks

Pros / Cons of this approach
• Surprisingly precise in cities
with larger numbers of
partitions
• Fast inference (<100ms on
t2.large)
• Excellent performance for
unique objects / landmarks
• Small model <300 MB can
be deployed anywhere
• Surprisingly inaccurate in
locales with fewer partitions.
• Poor performance for
common objects / terrain

AWS Infrastructure for WhereML

Architecture
1. Twitter Webhook calls to API Gateway endpoint
2. API Gateway invokes Lambda function with payload
from twitter
3. Lambda function calls out to SageMaker Inference
Endpoint with URL of image
4. Inference endpoint downloads image and classifies it
with LocationNet
5. Lambda posts results back to Twitter
AWS Lambda FunctionAmazon API
Gateway
SageMaker
Inference
Endpoint

Amazon API Gateway

AWS Lambda Function
• Proxy-Invocation from API Gateway sends entire request to the Lambda
• AWS Lambda Function:
1. Parses the incoming request
2. Verifies it is from Twitter and verifies message integrity
3. Parses the tweet
4. Sends the media URL in tweet to SageMaker endpoint
5. Uses twitter API to respond to original tweet
• Billed per GB/Second. 400,000 GB/S perpetual free-tier.
• Any scale of requests, but SageMaker endpoint limited to 10,000 TPS
• Python 🐍

Lambda Function – Verify Request
1 def lambda_handler(events, context):
2 # deal with bad requests
3 if event.get('path') != WEBHOOK_PATH:
4 return {'statusCode': 404, 'body': ''}
5 # deal with subscription calls
6 if event.get('httpMethod') == 'GET':
7 crc = event.get('queryStringParameters', {}).get('crc_token')
8 if not crc: return {'statusCode': 401, 'body': 'bad crc'}
9 return {'statusCode': 200, 'body': sign_crc(crc)}
10 # deal with bad crc
11 if not verify_request(event, context):
12 return {'statusCode': 400, 'body': 'bad crc'}

Lambda Function – Verify Request Utilities
1 def sign_crc(crc):
2 h = hmac.new(
3 bytes(CONSUMER_SECRET, 'ascii'), bytes(crc, 'ascii'),
4 digestmod=sha256)
5 return json.dumps({
6 "response_token": "sha256="+b64encode(h.digest()).decode()
7 })
8
9 def verify_request(event, context):
10 crc = event['headers']['X-Twitter-Webhooks-Signature']
11 h = hmac.new(
12 bytes(CONSUMER_SECRET, 'ascii'),
13 bytes(event['body'], 'utf-8'),
14 digestmod=sha256)
15 crc = b64decode(crc[7:]) # strip out the first 7 characters ("sha256=")
16 return hmac.compare_digest(h.digest(), crc)

Lambda Function – SageMaker and Twitter
1 def lambda_handler(events, context):
2 # we're good! load that event up
3 twitter_events = json.loads(event['body'])
4 for event in twitter_events.get('tweet_create_events', []):
5 if validate_record(event):
6 body = json.dumps({'url': media, 'max_predictions': MAX_PREDICTIONS})
7 results = json.loads(
8 sagemaker.invoke_endpoint(EndpointName=ENDPOINT_NAME, Body=body)
9 )['Body'].read()
10 status = build_tweet(results)
11 twitter_api.PostUpdate(
12 "📍 ?n" + status[0],
13 media=status[1],
14 in_reply_to_status_id=event['id_str'],
15 auto_populate_reply_metadata=True
16 )

SageMaker Architecture
• Docker container stored in ECR
• Autoscaling Endpoint spins up
containers as needed and
automatically fetches model artifacts
from S3 and puts them in
/opt/models
• Flask app responds to /ping and
/inference
Amazon SageMaker
Model Artifacts
Inference
Endpoint
Inference code
Amazon ECR
Inference code

SageMaker Inference Code – Load Model
1 import mxnet as mx
2 import numpy as np
3 sym, arg_params, aux_params = mx.model.load_checkpoint(MODEL_NAME, 12)
4 mod = mx.mod.Module(symbol=sym, context=mx.cpu())
5 mod.bind([('data', (1, 3, 224, 224))], for_training=False)
6 mod.set_params(arg_params, aux_params, allow_missing=True)
7 Batch = namedtuple('Batch', ['data'])
8 grids = load_grids()

SageMaker Inference Code – Predict
1 def predict(img, max_predictions):
2 mod.forward(Batch(img), is_train=False)
3 prob = mod.get_outputs()[0].asnumpy()[0]
4 pred = np.argsort(prob)[::-1]
5 result = []
6 for i in range(max_predictions):
7 pred_loc = grids[int(pred[i])]
8 result.append((pred_loc, prob[]))
9 return result
10
11 def download_and_predict(url, max_predictions=3):
12 img = preprocess_image(download_image(url))
13 return predict(img, max_predictions)

SageMaker Inference Code – Flask App
1 from flask import request, jsonify, Flask
2 import predict
3 app = Flask("WhereML")
4
5 @app.route("/ping")
6 def ping():
7 return "", 200
8
9 @app.route("/invocations", methods=["POST"])
10 def invoke():
11 data = request.get_json(force=True)
12 return jsonify(predict.download_and_predict(data['url']))

SageMaker Inference Code – Docker File
1 FROM mxnet/python:latest
2 WORKDIR /app
3 RUN pip install -U flask scikit-image numpy reverse_geocoder boto3
4 COPY *.py /app/
5 COPY grids.txt /app/
6 ENTRYPOINT ["python", "app.py"]
7 EXPOSE 8080

Working with Twitter

Twitter API
• Two ways to respond to mentions: User Streams API
(deprecated) and Account Activity API (beta)
• Account Activity Webhooks allow fully “serverless” approach
• UserStreams API requires running container/instance to poll
for updates
• UserStreams API is going away in June 2018… even though
the replacement for it is not GA…

Registering Webhook
1 twitter = OAuth1Session(**keys)
2 base_url = "https://api.twitter.com/1.1/all/env-beta/"
3 params={'url': "https://mywebsite.com/twitter/whereml"}
4 webhook_id = twitter.post(base_url+"webhooks.json", params).json()['id']
5 # pass webhook ID in prod, not needed in beta
6 twitter.post(base_url+"subscriptions.json")

SageMaker
Notebooks
Training
Algorithm
SageMaker
Training
Amazon ECR
Code Commit
Code Pipeline
SageMaker
Hosting
dataset
AWS
Lambda
API
Gateway
SageMaker Example End-to-End Architecture
Build
Train
Deploy
static website hosted on S3
Inference requests
Amazon S3
Amazon
Cloudfront
Web assets on
Cloudfront

Easy to get started!
Tons of tools
Machine Learning is FUN

Built live on twitch.tv/aws

Thank you!
randhunt@amazon.com

Where ml ai_heavy

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Similar to Where ml ai_heavy

Similar to Where ml ai_heavy (20)

More from Randall Hunt

More from Randall Hunt (9)

Recently uploaded

Recently uploaded (20)

Where ml ai_heavy

Editor's Notes