Skip to content
Success

Changes

Summary

  1. Clubb loss driver (#1303) (details)
  2. Clubb loss driver (#1303) (details)
Commit 24c744a9a68faba9b129d567379fd68cde6d1cda by noreply
Clubb loss driver (#1303)

* First set of commits for the new loss driver.

* Improvements

* Batching capabilities

* Lot's of cleanup, simplifications, and consolidation

* More testing

* Python interfaces for the loss driver

* Interface for tuner and tuner with modes.

* Nice updates

* Missed file

* Little updates to fix inconsistencies in the plots and averaging ranges

* Adding a script that converts sam and coamps netcdf files to a clubb named version, including things that need to be calculated. This fixes a bug where fields were being averaged before being calculated, see https://github.com/larson-group/clubb/issues/1284, but is currently only wired up in the new tuner and the dash app plotter

* Making some things public for nvhpc

* Organization and consistency test fix

* Adding taylor diagram to tune tab, this needs taylor metrics to be computed so the loss driver needed to output this info.

* Adding more fancy loss functions based on Taylor metrics and a way to switch how the loss is computed in the dash tuner

* Making NaN in code not an issue

* Mostly dash app improvements to plotting, also made subwindows for tuning

* Tuner tracking by case

* Removing stats time-variance code

* Simplifying loss driver output - now loss metrics are output explicitly rather than as packaged taylor_metrics and having different versions for subwindows and such

* Mainly ui updates and small simplifications

* Improving comments and cleaning things up a bit

* Removing strictness on stats window vs output interval, and adding a stats consistency test to tests for issues with batch size and output frequency and the windowing mode

* Fixing openmp bug and making incomplete stats intervals not output, as it was before

* Making jenkins test use standard stats, and fixing tiny bug that caused error running with debug+no rcp2 stats

* Adding gate around bindiff, there's an edge case where a cosmetic diff can trigger a full-diff even in a failed case, which should otherwise be considered BFB and skipped.

* Adding some README entries for the dash app and a python requirements file

* Adding comments and little bug fix to dash app

* Update to loss consistency test, making it simpler since stats consistency test covers most things. Also adding it to the stats consistency jenkins test since it is still similar

* Cleanup and code consolidation, new jenkins test for tuner

* Adding comments

* Fixing little bugs

* Slight improvement to stats_api

* Small bug fix

* Consolidating stats time setting
The file was modifiederror_code.F90 (diff)
The file was modifiedpdf_closure_module.F90 (diff)
The file was modifiedclubb_api_module.F90 (diff)
The file was modifiedstats_netcdf.F90 (diff)
Commit b6515b0aaca10ff234d8bc5f71659179640a3279 by noreply
Clubb loss driver (#1303)

* First set of commits for the new loss driver.

* Improvements

* Batching capabilities

* Lot's of cleanup, simplifications, and consolidation

* More testing

* Python interfaces for the loss driver

* Interface for tuner and tuner with modes.

* Nice updates

* Missed file

* Little updates to fix inconsistencies in the plots and averaging ranges

* Adding a script that converts sam and coamps netcdf files to a clubb named version, including things that need to be calculated. This fixes a bug where fields were being averaged before being calculated, see https://github.com/larson-group/clubb/issues/1284, but is currently only wired up in the new tuner and the dash app plotter

* Making some things public for nvhpc

* Organization and consistency test fix

* Adding taylor diagram to tune tab, this needs taylor metrics to be computed so the loss driver needed to output this info.

* Adding more fancy loss functions based on Taylor metrics and a way to switch how the loss is computed in the dash tuner

* Making NaN in code not an issue

* Mostly dash app improvements to plotting, also made subwindows for tuning

* Tuner tracking by case

* Removing stats time-variance code

* Simplifying loss driver output - now loss metrics are output explicitly rather than as packaged taylor_metrics and having different versions for subwindows and such

* Mainly ui updates and small simplifications

* Improving comments and cleaning things up a bit

* Removing strictness on stats window vs output interval, and adding a stats consistency test to tests for issues with batch size and output frequency and the windowing mode

* Fixing openmp bug and making incomplete stats intervals not output, as it was before

* Making jenkins test use standard stats, and fixing tiny bug that caused error running with debug+no rcp2 stats

* Adding gate around bindiff, there's an edge case where a cosmetic diff can trigger a full-diff even in a failed case, which should otherwise be considered BFB and skipped.

* Adding some README entries for the dash app and a python requirements file

* Adding comments and little bug fix to dash app

* Update to loss consistency test, making it simpler since stats consistency test covers most things. Also adding it to the stats consistency jenkins test since it is still similar

* Cleanup and code consolidation, new jenkins test for tuner

* Adding comments

* Fixing little bugs

* Slight improvement to stats_api

* Small bug fix

* Consolidating stats time setting
The file was modifiedlatin_hypercube_driver_module.F90 (diff)