Elm Quarter 3, 2017 Priorities

[PeterDSteinberg]

Elm Quarter 3, 2017 Priorities

These are the Phase II tasks most helpful to the preparation work we need to do for AGU December, and they are logically sequenced after the recent several months of work on Elm/xarray_filters/Earthio to separate ML from GIS.

• Milestone 2 Task 2: Hierarchical Modeling
• Milestone 2 Task 3: Vote Count Ensemble Averaging
• Milestone 6 Task 1: Support for xarray’s multi-file datasets
• Milestone 6 Task 2: Feature engineering options for 3-D and 4-D data

The following sections have a checklist of issues we should aim to finish before AGU December (the conference is the 11th to 15th week). Some of the listed issues are already completed or partially completed by PR #192 and related PRs in xarray_filters/Earthio

Elm

Docs / cleanup / promotion of Elm and related tools

• [ ] Doctesting throughout #206
• [ ] Documentation needs a flowchart overview of components #193
• [ ] Documentation reminder - Grib file support and relation to Python 2 vs 3 Grib #196
• [ ] Documentation issues under Xarray_filters Quarter 3, 2017 Priorities
• [ ] Any other issues labeled with "documentation" are reasonable but our time on example notebooks between now and December should be NLDAS oriented, not the various other example data sets mentioned on those issues
• [ ] Documentation issues under Earthio Quarter 3, 2017 Priorities

ML improvements

• [ ] Fix predict_many bugs related to package simplification #219
• [ ] Vote Count Ensemble Averaging #211
• [ ] Hierarchical modeling (fitting to predictions of already-fit estimators) #207
• [ ] Is avoid_repeated_params necessary? #203
• [ ] EaSearchCV - CV splitting is failing with Xarray inputs currently #204
• [ ] After new parameter selection in an evo algo, allow a cleanup/reject operation #205
• [x] Use the dill library instead of sklearn.externals.joblib to dump/load models #132
• [x] Improve testing of supervised learning and elm.pipeline.steps related-to-dask-xarray #171
• [ ] Ensure the hyperparameterization of pipelines allows model structure optimization #198
• [x] Refactor SklearnBase #197
• [x] Take advantage of new easy way to start dask-distributed related-to-dask-xarray #107
• [x] change the Pipeline.ensemble attribute to be a simple list not list of tuples #108
• [x] Remove client_context - just use Dask.Client ready related-to-dask-xarray #122
• [x] use dask-searchcv for evolutionary search EaSearchCV #192

Data structure improvements for ML related to this (see also this long term epic 202)

• [ ] Cross validation for xarray.* data structures #215
• [ ] Data structure / bug issues in Xarray_filters Quarter 3, 2017 Priorities
• [ ] Documentation issues under Earthio Quarter 3, 2017 Priorities

[PeterDSteinberg]

cc @gbrener @hsparra

[PeterDSteinberg]

Another task (assigned to me) for now until AGU is the improvement of Elm-Earthio-NLDAS. I will use that repo as a demo repo for discussion / sharing at AGU.