Squeezing Machine Learning into Serverless for Image Recognition - AWS Meetup Arlington VA
0 likes270 views
This document discusses using serverless computing to deploy machine learning models for image recognition. It describes automating the building of ML tools in a Lambda-compatible Docker container to package libraries and models. It then explains using the Serverless Framework to define the necessary infrastructure in AWS including an S3 bucket for image uploads, Lambda functions, and permissions to allow the functions to access S3 and perform image classification when a new image is uploaded. While this provides a starting point, it notes the solution is not as full-featured as using AWS Rekognition directly.
Squeezing Machine Learning into Serverless for Image Recognition - AWS Meetup Arlington VA
- 1. Squeezing Machine Learning into Serverless for Image Recognition Chris Shenton CTO at V! Studios AWS Meetup Arlington VA 2018-03-15
- 2. Shamelessly Plagiarized (or, “Inspired By a Blog Post”) ● Script builds ML libraries -- must be on AWS Linux AMI ● Create ZIP and upload to Lambda ● Manual creation of S3, roles, permissions, triggers But I’m way too lazy for that! https://blog.waya.ai/deploy-deep-machine-learning-in-production-the-pythonic-way-a17105f1540e
- 3. Automate Build of ML Tools in Lambda-Compatible Docker ● lambci docker image: compatible with AWS Lambda ● Makefile: ○ Avoid time-consuming builds of things we already have ○ Output binaries/libraries to Lambda function directory PIL Cython numpy torch torchvision Makefile Docker lambci/lambda Makefile (compile) lambdavision/ mounted by Docker lambdavision.py compiled ML libraries and binaries
- 4. Serverless Framework: Create Lambda, S3, Triggers, etc serverless.yml defines: ● S3 for image uploads ● Trigger to fire Lambda on S3 upload ● Permissions allowing Lambda to read S3 ● Lambda with Python function, ML libs ○ pulls pre-trained imagenet model with 1000 classes S3 Lambda lambdavision.py imagenet1000.py PIL/ cython/ numpy/ torch/ torchvision/ resnet18 pre-trained neural network upload trigger lambda lambdavision.py pulls trained model from network on cold start Logs score=11.207 classid=335 human=fox squirrel, eastern fox squirrel, Sciurus niger
- 5. It’s a Start, But It’s Not AWS Rekognition score=11.207 classid=335 human=fox squirrel, eastern fox squirrel, Sciurus niger score=9.719 classid=515 human=cowboy hat, ten-gallon hat score=13.650 classid=481 human=cassette ● Takes 3-8 seconds to classify a submitted image ● We hope to train a model on NASA-specific “stamps” ● Sensitive data my preclude use of AWS Rekognition ● Code: https://github.com/v-studios/LambdaVision
- 6. Want MOAR??
- 7. Makefile -- the heart of the build DSTDIR = lambdavision VENV = /tmp/venv3 PIP = $(VENV)/bin/pip PYTHON = $(VENV)/bin/python SITE_PACKAGES = $(VENV)/lib/python3.6/site-packages all: $(DSTDIR) docker run --rm -v "${PWD}":/var/task lambci/lambda:build-python3.6 make --debug lambda_packages
- 8. Makefile -- Inception -- run inside Docker lambda_packages: $(VENV) $(DSTDIR)/PIL $(DSTDIR)/Cython $(DSTDIR)/cython.py $(DSTDIR)/pyximport $(DSTDIR)/numpy-1.13.3-py3.6-linux-x86_64.egg $(DSTDIR)/torch $(DSTDIR)/torchvision $(VENV) $(PIP) $(PYTHON): virtualenv $(VENV) # Yaml is needed to build but not to run yaml: $(PIP) $(PIP) install pyyaml PIL: $(PIP) $(PIP) install Pillow cp -r $(SITE_PACKAGES)/PIL $(DSTDIR)/ Cython cython.py pyximport: $(PIP) $(PIP) install cython cp -r $(SITE_PACKAGES)/Cython $(DSTDIR)/ cp -r $(SITE_PACKAGES)/cython.py $(DSTDIR)/ cp -r $(SITE_PACKAGES)/pyximport $(DSTDIR)/
- 9. serverless.yml -- Define Infrastructure, Triggers, etc service: lambdavision provider: runtime: python3.6 memorySize: 2048 # Billed Duration: 8200 ms timeout: 30 iamRoleStatements: - Effect: Allow Resource: arn:aws:s3:::chris-lambdavision/* Action: - s3:GetObject functions: lambdavision: handler: lambdavision.s3upload events: - s3: bucket: chris-lambdavision event: s3:ObjectCreated:*
- 10. Live Demo?