🕘 Meeting notes: Machine Learning WG

May 2

  • Attendees
  • Joe Hamman / CarbonPlan
  • David John Gagne / NCAR
  • Zhonghua Zheng / NCAR (ASP)
  • Wei Ji Leong / Byrd Polar Climate Research Center
  • Max Jones / CarbonPlan

  • Agenda:
  • Joe: Finishing a climate downscaling process on Azure Cloud platform, mostly with scikit-downscale, deep learning using deepSD,
  • Max: Recently started at CarbonPlan a few weeks ago, getting acquainted with Pangeo/ML ecosystem, first developing on xbatcher, and then building loaders
  • David: Putting together real time ML storm mode (https://ncar.github.io/modeview) with js viewer, Python for data processing, AWS S3 for data storage
  • pyscript: runs Python directly in Javascript using pyodide/webassembly
  • Zhonghua: Developed tool for CESM … for urban climate, called Urban Climate Explorer. Next step is to do CMIP and …
  • xbatcher/torchgeo overlap
  • xbatcher designed to provide batches from a multidimensional dataset to feed into ML model
  • moving towards a data API to get data loaders from batch generators
  • roadmap includes ideas about random sampling
  • torchgeo currently using rasterio more as the backend, allows merging of datasets with different coordinate reference systems using Intersection/UnionDataset
  • David mentioned American Meteorological Society conference next Jan? Call for papers open. Could have a cloud machine learning oriented session, or something about open datasets? Have a tutorial like in AGU? Or host it independently.

April 4

  • Attendees
  • Joe Hamman / CarbonPlan
  • Zhonghua Zheng / NCAR (ASP)
  • Tom Augspurger
  • David John Gagne (NCAR)
  • Joe: Climate downscaling with lots of sklearn ML models and one DL model. Using prefect and dask, some scale-out issues. Cindy left but new person hired.
  • Douglas: 

February 7

  • Attendees
  • David John Gagne / NCAR 
  • Joe Hamman / CarbonPlan
  • Chris Dupuis / Columbia U. (formerly worked at GFDL)
  • John Schreck / NCAR