CoMET - Continental Model Engineering Toolkit
Created: 26 Apr 2023, 04:03 PM | Modified: =dateformat(this.file.mtime,"dd MMM yyyy, hh:mm a")
Tags: knowledge, courses, training
Dear AI & Data Community,
We’re happy to invite you to the next session of the AI & Data Community Meeting!
The upcoming meeting will provide an introduction into the planned “Integrated ML DevOps Tool Suite” which is based on the CoMET framework (see: https://github.geo.conti.de/Conti-heat-AI/comet).
The objective of this tool suite is to support, speed up and make “ML developement” more comfortable within Continental. The suite aims to cover the whole ML lifecycle.
This talk gives an introduction into the archtiectural concept of the tool suite, shows the current implementation status and provides a first example project.
ML devops tooling / comet more or less the same
https://confluence.auto.continental.cloud/display/projectAI/ML+DevOps+Tooling+-+Architectural+Design
Uses tensorflow TFX
Wandb planned to be hosted as an IT service soon
minIO - integrated storage for storing trained model / training data
How does this differ from AI Cluster by ADAS@BUD?
Data Science Orchestration Platform (DSOP)
- Hosted Kubeflow service
DSOP - training / more compute. remote docker installation
- Kubernetes cluster
Dsop interface using kubeflow
Wandb license - looking at deployment from IT side so that they can share
- AM alr has 30 → 42 licenses
Using comet for localization / cosma project
- Starting to refactor the
- They are using comet for training in cosma
Specify specific GPU instance?
- They are also not sure, they only know what is available on dsop
- Need to specify the memory requirements that you need - then they will select the instance for u
- Limitation from DSOP side
Currently DSOP is used by group IT
- But want to push it for everyone
Plans for this year 2023
- Generalizing the template / framework
Model deployment → during inference how do you do model monitoring for embedded