Post-processing for H models

[damonbayer]

Unfortunately, this has become a bit of monster PR, but it will result in a lot more consistency, clarity, and adaptability in the project.

Done

• Improves formatting of data files - no longer uses confusing names like "Disease" and "Other", instead uses more descriptive names like "observed_ed_visits" and "other_ed_visits" in a tidy format.
• Removes generation of legacy-formatted data files.
• Simplifies and improves output of timeseries models to more closely match PyRenew models
• Introduces group_time_index_to_date for robust PyRenew-index to date conversions.
• Introduces parse_pyrenew_model_name to extract expected features (h, e, w), based on the model's name
• Generalizes post-processing to work with models featuring any combination of h and e.
• Changes the offset argument in model scoring to be 1, rather than 1 / max_visits, since max_visits is less well-defined when working with multiple targets.
• There is now just a single scored.rds per model_run_dir. Different models and resolutions are indicated by their respective columns in these tables.
• Fixes a bug in collecting eval data for hospital admissions d10dcfb
• Mega epiweekly hubverse table is created, but I did not implement any post-processing functions for it.
• Simplifies post-processing to remove redundant data generation.
• Updates all parts of pipeline to work with the new formatted data files
• Major rewrite of plot collation script
• No R CMD check complaints
• Update hewr tests

Out of Scope

• Dashboard expects E models only.
  • Make dashboards work with arbitrary PyRenew-HEW models #311
• Notions of "daily" vs "epiweekly" data are not necessarily correct throughout. Generally, "daily" means "unaggregated" and "epiweekly" means "aggregated."
  • Therefore, the "epiweekly" hubverse table is more of an "aggregated" hubverse table, and there is not yet an "unaggregated" hubverse table.
  • Revise workflow to omit incorrect "daily" and "epiweekly" language. #330
• Some aspects of the batch post-processing should be re-worked as needed to accommodate some concept of a "preferred" forecast, which may be a mixture of different models for different locations. For now, we generate many disease_category_pointintervals plots, one for each model. We may wish to instead make a single plot for the "preferred" forecast.

Closes

Closes #308
Closes #296

[codecov]

Codecov Report

Attention: Patch coverage is 25.49505% with 301 lines in your changes missing coverage. Please review.

Project coverage is 24.47%. Comparing base (3a0a5bf) to head (792d3b1).
Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
hewr/R/process_state_forecast.R	5.00%	209 Missing ⚠️
pipelines/collate_plots.py	0.00%	34 Missing ⚠️
hewr/R/make_forecast_figure.R	0.00%	25 Missing ⚠️
hewr/R/directory_utils.R	15.38%	22 Missing ⚠️
pipelines/postprocess_forecast_batches.py	0.00%	7 Missing ⚠️
pipelines/prep_data.py	0.00%	4 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #287      +/-   ##
==========================================
- Coverage   25.56%   24.47%   -1.10%     
==========================================
  Files          22       22              
  Lines        1682     1704      +22     
==========================================
- Hits          430      417      -13     
- Misses       1252     1287      +35     

Flag	Coverage Δ
hewr	41.68% <28.69%> (-7.94%)	⬇️
pipelines	4.67% <0.00%> (+0.32%)	⬆️
pyrenew_hew	27.96% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[dylanhmorris]

Thanks for this, @damonbayer. All my comments are addressed. Let me know when remaining test issues are fixed.

[dylanhmorris]

Going to try an end-to-end run now.

[dylanhmorris]

Fitting and postprocessing run end to end! Thank you, @damonbayer.

Two things before merge:

1. Afaict everything looks as expected, except that the titles on the prop_disease_ed_visits plots are wrong. They say "Other Emergency Department Visits". Presumably a bug in the figure labeling logic somewhere. I think this is worth fixing before merge, as it should be a quick fix.
2. The mega hubverse table is big enough that I don't think it's going to be readily human-readable, and any submitted hubverse tables will be downsampled from it. Given that, worth saving it as parquet rather than .tsv?

[damonbayer]

@dylanhmorris I've addressed both of your most recent comments.

[dylanhmorris]

LGTM. Thanks for all your hard work, @damonbayer!