Loading...

​​ML WG #3

​​Attendance

• ​​Noah

• ​​David

• ​​Stephan

• ​​Tom

​​Agenda

• ​​Progress updates

​​

​​Notes

• ​​Distributed training: horovod

• ​​Dask communications layer using nvlink- blog article in progress: https://github.com/dask/dask-blog/pull/28

• ​​PyTorch & distributed discussion https://github.com/dask/distributed/issues/2581

​​

• ​​Noah. Workflow

• ​​cloud training

• ​​stephan: use GPU at paperspace, and AWS (Several environments)

• ​​environments with tensorflow is annoying

• ​​Tom: conda should improve this

• ​​Need for intermediate models

• ​​Hard-coded NN in fortran code: (hard to change)

• ​​Online learning. Python climate model interface

• ​​Streamz being maintained at anaconda (martin)

• ​​would interface with Kafka (and other data sources)

• ​​provide higher level API

• ​​How suitable for deep learning pipelines

• ​​Data loading issue

• ​​Summarize this https://github.com/pangeo-data/pangeo/issues/567 ​@Noah B​ 

• ​​https://github.com/dmlc/dlpack

• ​​cuda__array_interface. no-copy operations. could be good for evaluation in climate model

• ​​Not so useful for on-disk?

• ​​Blog post. loop in Tom, Stephan, and Ryan.

• ​​nice benchmarks for data loading

• ​​toy problem of data loading. add some simple ML model. time ratio of training to loading

• ​​need to detect bottlenecks (check hardware).

• ​​Next meeting. Monday July 1 9AM

​​