MXNet Serverless example
Running MXNet container on lambda to create a streaming inference pipeline.
This is based on the following article: https://medium.com/apache-mxnet/streaming-inference-pipeline-deploying-mxnet-model-on-aws-lambda-7ce6bc8f4cc8
Running
Create the initial resources:
make setup
Upload the model's resources into the resource bucket:
make upload-resource
To make a prediction, assuming your images are in examples sub-dir:
make predictions
To view the results of the predictions, download the content of the target bucket:
make download-output
Docker compatibility
To enable your lambda function to run within a docker container in lambda, your function need to be run by the AWS Lambda RIC:
https://github.com/aws/aws-lambda-python-runtime-interface-client
The lambda function is created via a docker container containing the AWS Lambda python RIC.
A multi stage build whereby we build the RIC client with the app dependencies and package them into a function directory together with the app files
The function directory is then copied into the final image which contains the ML lib in this case MXNet you want to run the lambda function on.
The final image needs to be pushed to ECR as within the lambda console you can only select images hosted on ECR.
Testing it locally
Left as exercise to reader:
https://github.com/aws/aws-lambda-runtime-interface-emulator
License
Licensed under an MIT license.