Results and I/O¶

`ForwardResult` outputs¶

Two CSV writers persist forward-model products:

from sed_model import save_sed, save_magnitudes

save_sed(result, "forward_sed.csv")            # wavelength, surface_flux, observed_flux
save_magnitudes(result, "forward_magnitudes.csv")  # filter, magnitude, band_flux + metadata

Both files carry a #-commented header recording the input parameters (teff, logg, meta, R, d, mag_system, ...) so the file is self-describing.

`InverseResult`¶

Returned by run_inverse; a frozen container with:

Field	Description
`samples`	`(n_samples, n_free)` flattened post-burn-in chain; columns follow `param_names`
`log_prob`	`(n_samples,)` log-posterior per sample
`param_names`	the free parameters, canonical order (e.g. `['teff', 'logg', 'meta']`)
`fixed_params`	`{name: value}` for the parameters that were held fixed
`obs_magnitudes`, `obs_uncertainties`, `filter_names`	the data that were fitted
`R`, `d`, `mag_system`	physical configuration
`n_walkers`, `n_steps`, `n_burn`, `n_thin`	sampler settings
`acceptance_fraction`	per-walker acceptance fractions
`autocorr_time`	integrated autocorrelation time per free parameter, or `None`

Summaries¶

s = posterior.summary()
# {'teff': {'median': ..., 'lo': ..., 'hi': ...,
#           'lower_1sigma': ..., 'upper_1sigma': ...}, ...}

lo/hi are the 15.865 / 84.135 percentiles (±1σ equivalent); lower_1sigma/upper_1sigma are the distances from the median to those percentiles.

posterior.print_summary()

prints a formatted block including the sampled and fixed parameters, the median ± asymmetric 1σ for each free parameter, mean acceptance fraction, and autocorrelation times.

MAP estimate¶

map_params = posterior.map_estimate()
# {'teff': ..., 'logg': ..., 'meta': ...}  — the sample with highest log_prob

Returns a dict keyed by parameter name, so it works for any free-parameter configuration (3 parameters or 5).

Persistence¶

posterior.save("posterior.npz")          # .npz appended if missing
loaded = InverseResult.load("posterior.npz")

The .npz stores the chain arrays plus a JSON-encoded metadata blob (observations, parameter names, fixed values, sampler settings), so a loaded result round-trips exactly — summary(), map_estimate(), and print_summary() all work on the reloaded object.