Serverless Computer Vision

This repo is an attempt at setting up serverless computer vision applications using AWS services. For this implementation I have used TensorFlow but the stack can be used with any application/framework.

You can leverage this package and its packages to build your own pipeline using AWS.

Frontend

• React
• Redux

Backend

• S3
• AWS Lambda
• API Gateway
• Cognito

Deployment & Testing

• AWS Cloudformation
• Docker
• aws-sam-local

Stack Files

• cfn-template.yml - Cloudformation template to setup necessary S3 buckets, Cognito Authentication, IAM roles and policies
• template.yml - Contains SAM template for Lambda function and API Gateway
• swagger.yml - Contains API specifications
• run.sh - Bash script containing all the necessary commands to setup and deploy the stack.

Requirements

• nodejs
• Python
• pip
• viratulenv

Setup Instructions

1. Setup virtualenv and install node_modules

    bash run.sh setup

2. Create IAM user on AWS with admin access and use the access_key and secret_key to configure aws credentials locally.

    bash run.sh aws_config

3. Copy/Download the model(s) into /models directory. You can use my Yolo model for this setup

    wget https://www.dropbox.com/s/h8ywy9lp8siw0ml/yolo_tf.pb -P models/

4. Deploy the entire stack

    bash run.sh deploy

OR

4. Deploy the basic stack containing S3, Cognito and IAM.

    bash run.sh deploy_stack

5. Upload model(s) to S3 bucket.

    bash run.sh deploy_model

6. Build and deploy static assest i.e entire frontend built using React

    bash run.sh deploy_website

7. Build and deploy lambdas

    bash run.sh deploy_lambdas

Development & Testing

• Follow the first 4 steps from setup instructions.

• Start the api locally

    npm run sam local start-api

• Replace API_URL in /src/actions/index.jsx with the local url.

• Run the local node server locally.

    npm run watch

• Open http://localhost:3000/ in the browser

• Drop images in the application to upload them to S3.

• If you want to debug the function you can invoke the function locally once the images are in S3. The will generate better logs.

    npm run sam local invoke ProcessImage -- --event fixtures/ProcessImage.json

Webpage

Lambda Limits

• Code Size: 261 Mb
• Persistant Storage: 512 Mb