ML model deployment using Serverless Framework

Deploying Machine Learning models on AWS using Serverless Framework.

In this repository, we'll be building the following architecture:

Don't forget to set up your AWS credentials using aws cli.

Details

The input data will be stored in an S3 bucket, in our case, the csv file.

The model file will also be stored inside a S3 bucket. Since the model file can be large (>90Mb), we must load it during AWS Lambda inference execution from Amazon S3. You can observe that the model downloading was placed outside of the handler function. This was done to take advantage of AWS Lambda container reuse. Any code executed outside the handler method will be invoked only once upon container creation and kept in memory across calls to the same Lambda container, making subsequent calls to Lambda faster.

Files

• s3_utils.py: functions for read and write from s3
• config.py: load environments variables
• serverless.yml: Serverless Framework YML file
• handler.py: file to be invoked for AWS lambda when the service executes the code.
• requirement.txt: python dependencies
• .env: environment variables
• Images/: contains all images from repository

Usage

Installing Serverless Framework

npm install -g serverless

In this repository will be using Python3. Create a repository, go inside and run the following command:

serverless create --template aws-python3

Installing serverless-python-requirements plugin. A Serverless v1.x plugin to automatically bundle dependencies from requirements.txt and make them available in your PYTHONPATH.:

sls plugin install -n serverless-python-requirements

Installing serverless-dotenv-plugin. Preload environment variables into serverless. Use this plugin if you have variables stored in a .env file that you want loaded into your serverless yaml config.:

npm i -D serverless-dotenv-plugin

Deployment

serverless deploy

Code

The model used is an energy forecasting model using XGboost. The dataset consists of hourly energy consumption rates in kWh for an industrial utility. You can check more about it in here.

The model has a forecasting horizon (The number of time periods to forecast into the future) of 1 hour.

The model inputs are stored in s3 Bucket. It consists of the model inputs parameter of the last hour. The parameters are:

• hour
• weekday
• dayofyear
• is_weekend
• weekofyear
• month
• season
• shift
• lag_1
• lag_8
• lag_25

Having energy forecast as output.

Inputs

The inputs are saved in an s3 bucket as a .csv

Model

The XGboost model is saved as .sav in s3

Results

The results are also stored in s3 as a .csv

Pipeline

CloudWatch trigger (every hour) -> Read Inputs -> Read Model -> Prediction -> Save Model

Testing function locally

serverless invoke local -f function_name