Skip to content

This repository contains the support of a talk about ML in production (knative, ONNX, and Go)

Notifications You must be signed in to change notification settings

akashrv/khappygo

 
 

Repository files navigation

khappygo

This repository contains the support of a talk about ML in production (knative, ONNX, and Go)

This repository contains all the code to:

  • instantiate a k8s cluster on GCP (via terraform)
  • add the knative eventing CRDs (channel is InMemory by now)
  • add the tracing system with zipkin (again in memory)
  • create the triggers
  • compile and deploy the services

Services

The goal of the application is to create a database of sentiments of people from a picture. The application is composed of several services:

  • face detection
  • image extraction
  • sentiment analysis
  • data management
  • ui

The face detection and sentiment analysis are deep neural networks pre-encoded with ONNX.

Note: the code is not fully tested, and probably only works in my environment.

Goals

The goal of this presentation are multiple:

  • to show that a ML service can be decoupled (ports and adapters): The business logic (software 2.0) is using ONNX, and the regular part is using software 1.0. (see this blog post for more info)
  • to show how to pass from a program to a software with "domain events" (see this talk for the différence)
  • to show how knative eventing is a good support for the services
  • to show how a service can easily replaced by changing the trigger (passing from "yolo" to "pigo" - aka: you don't need deep learning :))

But most of all:

  • Dev Ops and Data-Scientists mustcollaborate, but, thanks to a good level of abstraction, they don't need to master each-other's area of expertise.

Workflow

This is a simple (and outdated) workflow, describing the basic principle:

Result

Source

Result

Traces

Installation

TODO

Models

Deep learning

The deep learning models are encoded in ONNX and stored using git-lfs. see the gofaces repo for more info.

Two pre-trained models are used:

  • Tiny Yolo v2
  • emotion fer+

non deep learning

This POC also uses pigo as an improvement to the face detection service. The cascade file is also included in this repo.

Slides

The slides are in progress. They are in the present format an best viewed with github.com/owulveryck/present

About

This repository contains the support of a talk about ML in production (knative, ONNX, and Go)

Resources

Stars

Watchers

Forks

Packages

No packages published

Languages

  • Go 42.9%
  • Shell 26.6%
  • Smarty 15.1%
  • Makefile 5.2%
  • JavaScript 4.2%
  • CSS 3.1%
  • Other 2.9%