serverless-sagemaker-groundtruth

This serverless plugin includes a set of utilities to implement custom workflow for AWS Sagemaker Groundtruth

Currently includes :

Serve liquid template from manifest file + prelambda the same it is done on AWS Sagemaker Groundtruth
Run End to end test pre-lambda -> labelling simulation -> post lambda

Any Pull request will be warmly welcome !

Ideas for future implementation :

Create Tasks from serverless CLI
Test Chained tasks
Expose nodejs api to integrate with testing suites

Installation

npm install --save-dev serverless-sagemaker-groundtruth

Usage as a serverless plugin

Example serverless.yml

In order to use this module, you need to add a groundtruthTasks key into your serverless.yml file

...

plugins:
  - serverless-sagemaker-groundtruth

functions:
  pre-example: 
    handler: handler.pre
    name: pre
  post-example: 
    handler: handler.postObjectDetection
    name: post

groundtruthTasks:
  basic:
    pre: pre-example
    post: post-example
    template: app/templates/object-detection/basic.liquid.html

Serve a liquid template against a manifest file

serverless groundtruth serve \
  --groundtruthTask <groundtruthTask-name> \
  --manifest <s3-uri or local file> \
  --row <row index>

Test e2e behavior of sagemaker groundtruth workflow

The puppeteer module example

Here, we create a puppeteer module which is doing random bounding boxes (using hasard library) :

const BbPromise = require('bluebird')
const h = require('hasard');

/**
* This function is binding a sequence of actions made by the user before submitting the form
* This is an example showing how to simulate a use bounding box actions
* @param {Page} page puppeteer page instance see https://github.com/puppeteer/puppeteer
* This page is open and running in the annotation page
* @param {Object} manifestRow the object from the manifest file row
* @param {Object} prelambdaOutput the output object from the prelambda result
* @returns {Promise} the promise is resolved once the user has done all needed actions on the form
*/

module.exports = function({
	page, 
	manifestRow,  
	workerId
}){
	
	// we draw 5 boxes for each worker
	const nBoxes = 5;
	
	// Cat and Dog
	const nCategories = 2;
	
	// Using the technic from https://github.com/puppeteer/puppeteer/issues/858#issuecomment-438540596 to select the node
	return page.evaluateHandle(`document.querySelector("body > crowd-form > form > crowd-bounding-box").shadowRoot.querySelector("#annotation-area-container > div > div > div")`)
		.then(imageCanvas => {
			return imageCanvas.boundingBox()
		}).then(boundingBox => {
			
			// define a random bounding box over the image canvas using hasard library
			// see more example in https://www.npmjs.com/package/hasard
			const width = h.reference(h.integer(0, Math.floor(boundingBox.width)));
			const height = h.reference(h.integer(0, Math.floor(boundingBox.height)));
			const top = h.add(h.integer(0, h.substract(Math.floor(boundingBox.width), width)), Math.floor(boundingBox.x));
			const left = h.add(h.integer(0, h.substract(Math.floor(boundingBox.height), height)), Math.floor(boundingBox.y));

			const randomAnnotation = h.object({
				box: h.array([
					top,
					left,
					width,
					height
				]),
				category: h.integer(0, nCategories-1)
			});
			
			const workerAnnotations = randomAnnotation.run(nBoxes)
			
			return BbPromise.map(workerAnnotations, ({box, category}) => {
				return page.evaluateHandle(`document.querySelector("body > crowd-form > form > crowd-bounding-box").shadowRoot.querySelector("#react-mount-point > div > div > awsui-app-layout > div > div.awsui-app-layout__tools.awsui-app-layout--open > aside > div > span > div > div.label-pane-content > div:nth-child(${category+1})")`)
					.then(categoryButton => categoryButton.click())
					.then(() => page.mouse.move(box[0], box[1]))
					.then(() => page.mouse.down())
					.then(() => page.mouse.move(box[0]+box[2], box[1]+box[3]))
					.then(() => page.mouse.up());
				
			}, {concurrency: 1})
		}).then(() => {
			console.log(`${workerId} actions simulation done on ${JSON.stringify(manifestRow)}`)
			// at the end we return nothing, serverless-sagemaker-groundtruth will automatically request the output from the page
		})
}

The end to end command

serverless groundtruth test e2e \
	--groundtruthTask <groundtruthTask-name> \
	--manifest <s3-uri or local file> \
	--puppeteerModule <path to the module> \
	--workerIds a,b,c

Usage programmatically

You can use serverless-sagemaker-groundtruth functions in your nodejs code by using

const gtLibs = require('serverless-sagemaker-groundtruth/lib')

endToEnd


/**
* @param {String} template path to the liquid template file
* @param {String} labelAttributeName labelAttributeName to use as output of the postLambda function
* @param {Object} manifestRow js object reproesnting the manifest row
* @param {Function} preLambda js function to use as pre lambda function
* @param {Number} [port=3000]  port to use to serve the web page
* @param {Function} postLambda js function to use as post lambda function
* @param {Array.<String>} workerIds js function to use as post lambda function
* @param {PuppeteerModule} puppeteerMod module that simulate the behavior of a worker
* @returns {Promise.<PostLambdaOutput>}
*/

return gtLibs.endToEnd({
	template,
	labelAttributeName, 
	manifestRow,
	preLambda,
	port,
	postLambda,
	workerIds,
	puppeteerMod
});

Remarks

Local consolidation request file compatibilty

You need to make sure that you post lambda function is compatible with using local filename in event.payload.s3Uri. You can use gtLibs.loadFile if you need such a function

Template

Your template should be submited using a button that can match with button.awsui-button[type="submit"] selector.

Serverless Sagemaker Groundtruth