Open discussions on specific topics selected by the Software Working Group and selected from the list of SWG Topics For Discussion.
Round Table Discussion: ML Flow, moderated by Ben Galewsky.
Want to learn how to make ML Training more productive and reproducible? Want to share trained models within your group and then publish them as citable objects? Ben will be showing
an open source tool for machine learning Ops, MLFlow and how it can be used in research. He will demonstrate the slurm plugin for launching training runs on Delta and provide
concrete examples from the Quantum Monte Carlo project he is working on.
Recording: https://uofi.box.com/s/n7fftnyeh76wlgvd2kaj2vn2hvbaigp2
Slides:
Attendees: Ben Galewsky, Max, Burnette, Doug Friedel, Bingji Gu, Dipa Dey, Mark Neubauer, Minu Mathew, Rob Kooper, Luigi Marini, Jim Phillips, Bing Zhang, Jim Gaglione (guest), Jonathan Kim, Marcos Frenkel, Matt Berry, Nick Kowalik, Pengyin Shan, Natalie Romer (guest), Rebecca Eveland, Santiago Nunez-Corrales, Sara Lambert, Sonali Joshi (guest), Ya-Lan Yang, Tim McFadden (guest) Visu Mohanarajan, Chris Navarro, Sandeep Puthanveetil Satheesan, Kastan Day, Lisa Yanello
Ben discussed what MLflow tracking does for reproducible and Citable training. Anyone with HAL can launch an ML Flow in a matter of minutes. There are many models to choose from.
Discussion of models concepts with packaging, dependencies, creation utilities, etc.
Bingji notes that cookie cutters could work in ML flow.
There is a check box to ask ML flow to enable, but you don't always want to do this.
Mark notes that scalability is not an issue because we have a server set up for each individual/group.
Rob suggests setting up Mark's request to use this through Illinois Computes.
Sandeep asked about a project file, and how parameters change. Ben notes you need to send parameters in postgres through ML flow, including hyper parameters.
Comments
Links mentioned in this Round Table:
Best Practices Handbook: https://github.com/ncsa/software-development-handbook
If you are interested in contributing to a Round Table, please see these links:
Round Table Google Sheet: https://docs.google.com/spreadsheets/d/1kbgO6sIb_4eLugfSVKQNCTXdaKp1R6m0RDczPTsUAoQ/edit#gid=0 Every one should have edit permission.