specsrbench.build

Rebuilding the cache the figures read, from JADES DR4 and the Hub.

Six stages, each consuming what the one before it wrote:

predictions   paired dataset + Hub checkpoints -> ML predictions
sets          predictions -> eval / calib / tune sets, galaxy-disjoint
lsf           raw JADES x1d + line fits       -> the measured LSF kernel
tune          tune set + kernel               -> classical_params.json
classical     eval set + kernel + parameters  -> the classical caches
lines         every reconstruction            -> Gaussian fits, S/N, summary

Only lsf needs the raw JADES tree, and only predictions needs torch and network access to the Hub; the rest run on what the earlier stages wrote.

The order matters in a way that is easy to get wrong. lsf reads the line fits that lines writes, so the very first build of a fresh tree runs lines once against the shipped kernel, then lsf, then tune -> classical -> lines again on the measured one. classical refuses to write a cache built with the shipped kernel, which is what stops that first pass being mistaken for a finished one.

Module Attributes

STAGES

stage name -> module implementing it.

Functions

require_npz(path, produced_by)

Load an input, or say which stage produces it.

run_stage(stage[, args])

Run one build stage.
specsrbench.build.STAGES: dict[str, str] = {'classical': 'classical_cache', 'lines': 'lines', 'lsf': 'lsf', 'predictions': 'predictions', 'sets': 'sets', 'tune': 'tune'}

stage name -> module implementing it.

specsrbench.build.run_stage(stage, args=None)[source]

Run one build stage. args is the parsed CLI namespace, if any.

Parameters:

stage (str)

Return type:

int

specsrbench.build.require_npz(path, produced_by)[source]

Load an input, or say which stage produces it.

A build stage that cannot find its input is the normal state of a fresh clone, not a corruption. Saying so, and naming the command that fixes it, is the difference between a one-line fix and an afternoon spent looking for a cache that was never there.

Parameters: