Add fitting infrastructure

[ethho]

• Add infrastructure for fitting an Episimlab model to hospital line list data, estimating a global transmission parameter like beta as the dependent variable.
• As a first pass, implement linear least squares fitting (scipy.optimize.least_squares) similar to SEIR-city v1.4
  • Most of the v1.4 implementation is data structure munging, so the Episimlab implementation should be much more lightweight in general.
• All fitting infra will live in the new episimlab.fit submodule, with respective pytests at tests/test_fit.

[ethho]

@kellypierce refresh my memory: I believe we want a loss function similar to calc_residual. This function returns a 1-D vector of (the residual error of) Ih over the time dimension (every time point in the SEIR simulation), correct?

[kellypierce]

@kellypierce refresh my memory: I believe we want a loss function similar to calc_residual. This function returns a 1-D vector of (the residual error of) Ih over the time dimension (every time point in the SEIR simulation), correct?

Yes, that is correct.

[ethho]

To Do

• [x] llsq uses attr class instead of scrappy functions
• [x] Munge compartments Ih and Ih2D from line list data data/CV19Hospital_ICU_DeID_20210414.csv, generating schema similar to:

  date,zip_code,hosp,deceased

  • Done on branch munge-ll-hosp (merged below)
  • Adds examples/20210504_munge_ll_hosp_data.ipynb, which reads input file data/CV19Hospital_ICU_DeID_20210414.csv and writes an xarray.DataArray to data/CV19Hospital_ICU_DeID_20210414.nc implemented on 98fe51c but since removed
    • Forgot to perform += operation on previous row.
    • This is probably not easy to do.
    • As of d67b713, writes the delta series instead, to data/CV19Hospital_ICU_DeID_20210414_delta.csv, it's much smaller anyway.
• [ ] Try fitting to munged hosp data (data/CV19Hospital_ICU_DeID_20210414_delta.csv)

[ethho]

Takeaways from Ethan-KP Sync

• Simply ignore rows with a null discharge date
• xslx schema is relatively stable, but directly from data provider (not preproc pipeline), so potentially prone to change
  • Doesn't really matter, since we might want to start fitting to ICU instead of head in beds in the near future

[ethho]

As of ef9a908, fitting to heads in beds summed over all vertices (zip codes).