Apr 29, 2019

Attendees

@Ryan A / Columbia, LDEO / rabernat

Noah Brenowitz / UW / nbren12

Joe Hamman

Jim Bednar / Anaconda / jbednar

David Gagne / UCAR /

Agenda

Ryan’s local stencil example

Noah: Sampler/SklearnWrapper

tfRecords vs PyTables

DJ: https://github.com/djgagne/ml-workflow-examples/blob/master/djgagne_partial_dependence_plot.ipynb

Discussion

Ryan’s example

Learn Geostrophic balance on lat/lon grid

CNN didn’t work

Augment data with 3d coordinates (x,y,z), struggled with periodic BC

XBatcher

Many samples (millions)

Jim

As part of new NASA/Anaconda project, wanting to collect workflows to optimize as best practices. Will be monitoring the ml-workflow-examples and adapting them if appropriate.

David

Interpretability, partial dependence

Super cells

U,V,T → 1 km vorticity

csv data (84 MB)

model agnostic

Questions:

How will dask do on bigger problems?

Could be made data parallel by changing order of “means”

Ryan: tensorflow didn’t like dask-based DataLoaders

Action items

Let’s get this done @someone

​​Apr 29, 2019

​​Attendees

​​Agenda

​​Discussion

​​Ryan’s example

​​Jim

​​David

​​Action items