DEV Community

Matt Coulter for AWS Heroes

Posted on

Deploying a ML model using the new AWS Lambda Container Image Functionality

This week at re:Invent we saw AWS announce the ability to bring your own container to Lambda functions. This may seem very anti-serverless at first but remember that serverless is a spectrum and everyone is in a different place.

The scenario that gets me super excited is the deployment of ML models inside Lambda. Before, it was just too awkward for most use cases to do ML in Lambda but I will demonstrate below that it is very easy now. To do so I will use a new CDK Pattern

The Predictive Lambda Pattern


This is a pattern that uses a container inside Lambda to deploy a custom Python ML model to predict the nearest Chipotle restaurant based on your lat/long.

Some Useful References:

Available Versions

What's Included In This Pattern?

This pattern uses sklearn to create a custom k nearest neighbour model to predict the nearest Chipotle to a given Latitude and Longitude. The model is deployed inside a container attached to AWS Lambda.

The Data

If you want to look at the data used for this model you can look at the jupyter notebook, the raw data came from kaggle

The ML Model

This is a very simple model to demonstrate the concept (I didn't even check the accuracy because it doesn't change the pattern). It uses sklearn nearest neighbors to predict the closest Chipotle location to a given lat/long

Two Docker Containers

I use the Lambda image to train the ML model in one container and then I use a separate container for the deployed Lambda Function. The reason I do this is because it means that you know you have pickled your model in the same environment it will be deployed but you can use things that wont be packaged into your deployed function keeping it as lightweight as possible. You will also have a built container image containing the raw data, the training logic and the trained model. These images could be archived to have a history of your model.

A Lambda Function

I have this setup with a 15 second timeout and 4GB ram to comfortably run our model


Setup as a proxy integration, all requests hit the Lambda Function

How Do I Test This Pattern?

do "npm run deploy" from the base directory and you will have the url for an API Gateway output into the logs or in the CloudFormation console. Open that url in a browser but add "?lat=39.153198&long=-77.066176" to the end and you should get back a prediction.

Deep Dive WalkThrough

There are 3 separate processes included in this pattern

  1. A scripted process to train and export a ML model from inside the Lambda Python image for runtime compatibility
  2. A Dockerfile to take that exported model and use it inside a containerised lambda function
  3. A CDK implementation to deploy an API Gateway and the above Lambda

Model Training - Completely Optional

I have included a pre-trained model in this pattern so you only need to do this if you want to understand how I did it or you want to try it with your own model.

If you look inside the model folder there is a shell script called, running this script (making sure you have docker started) will completely retrain the model.

cd model
Enter fullscreen mode Exit fullscreen mode

The code in this shell script looks worse than it is

#Using the named TrainingDockerfile, build this image and tag it as chipotle
docker build . -f TrainingDockerfile -t chipotle
#We need the image id, so query docker for an image tagged with chipotle
IMAGE_ID=$(docker images -q chipotle)
#Start the image as a background process named training
docker run -d --name 'training' ${IMAGE_ID} 'app.handler'
#Copy the trained model out of the container
docker cp training:/var/task/chipotle.pkl chipotle.pkl
#stop the running instance and delete it
docker kill training && docker rm training
Enter fullscreen mode Exit fullscreen mode

The next place to look is TrainingDockerfile

# Use the python lambda image from AWS ECR
# Put these 3 files inside the container
COPY training/ requirements.txt training/chipotle_stores.csv ./
# Install python dependencies
RUN pip3 install -r requirements.txt
# Run the training logic
RUN python3
Enter fullscreen mode Exit fullscreen mode

If you want to look at the data inside chipotle_stores.csv you can look at the jupyter notebook, the raw data came from kaggle

The training logic inside training/ loads chipotle_stores.csv into Python, cleans it up and then trains/exports a model. The training/export logic is

#train model
model = KNeighborsClassifier(n_neighbors=2, weights="distance", algorithm="auto"), train_set_labels)

#export model
joblib.dump(model, 'chipotle.pkl')
Enter fullscreen mode Exit fullscreen mode

Containerised Lambda Function

Most of the logic to make this happen is in model/Dockerfile

# copy our function logic, requirements and model into the container
COPY deployment/ requirements.txt chipotle.pkl ./
# install the dependencies
RUN pip3 install -r requirements.txt
# the lambda handler is located inside as a method called lambdaHandler
CMD ["app.lambdaHandler"]
Enter fullscreen mode Exit fullscreen mode

The actual lambda handler code inside deployment/ is the same as any other lambda function

import joblib

def lambdaHandler(event, context):
    model = joblib.load('chipotle.pkl')

        latitude = event["queryStringParameters"]['lat']
    except KeyError:
        latitude = 0

        longitude = event["queryStringParameters"]['long']
    except KeyError:
        longitude = 0

    prediction = model.predict([[latitude,longitude]])
    prediction = prediction.tolist()
    return {'body': str(prediction[0]), 'statusCode': 200}
Enter fullscreen mode Exit fullscreen mode

CDK Infra Logic

The relevant piece of CDK is that instead of the normal way of creating our function, we use lambda.DockerImageFunction and ask CDK to build our container from the model folder

// defines an AWS Lambda resource
const predictiveLambda = new lambda.DockerImageFunction(this, 'PredictiveLambda', {
    code: lambda.DockerImageCode.fromImageAsset(path.join(__dirname, '../model')),
    timeout: cdk.Duration.seconds(15)
Enter fullscreen mode Exit fullscreen mode

Top comments (0)