specsrbench.build.sets

Stage 2 – cut the predictions into the three galaxy-disjoint sets.

Everything downstream reads one of these three, and which one is not a detail:

eval_set.npz

572 held-out originals. Every published number.

tune_set.npz

40 galaxies from the training side. The classical parameters are chosen on this and nowhere else.

calib_set.npz

400 galaxies, also from the training side. Guard 4 – whether a resolvable [O III] pair survives a filter – is measured on these.

The two smaller sets exist because tuning a method on the set you then report it on is the same error, one level up, that the group-wise split fixes at the galaxy level. All three draw from originals only: the paired dataset carries 21 augmented rows per galaxy, and a set built from augmented rows would measure how well a method deconvolves a copy of a spectrum it has already seen.

Guard 4 is measured on calib_set rather than on a synthetic pair on purpose. A synthetic doublet was tried and discarded: it is far easier to hold apart than a real one, and it passed settings that merge in real data.

Nothing here is random in the sense of being irreproducible – the draws are seeded and the seeds are part of the definition of the sets.

Module Attributes

TUNE_N	Sizes and seeds of the two training-side draws.
TUNE_SEED	Sizes and seeds of the two training-side draws.

Functions

build_eval_set(predictions, out)	eval_set.npz -- the predictions plus the normalised spectra.
main([argv])	Cut eval / tune / calib sets from the predictions and the paired dataset.
training_originals(dataset)	Row indices of the un-augmented galaxies on the training side.
zscore(a)	Per-spectrum z-score, plus the moments needed to undo it.

specsrbench.build.sets.TUNE_N = 40

Sizes and seeds of the two training-side draws. These are the definition of those sets: changing either changes which galaxies the classical methods were tuned and guarded on, and therefore every classical number downstream.

specsrbench.build.sets.TUNE_SEED = 0

Sizes and seeds of the two training-side draws. These are the definition of those sets: changing either changes which galaxies the classical methods were tuned and guarded on, and therefore every classical number downstream.

specsrbench.build.sets.zscore(a)

Per-spectrum z-score, plus the moments needed to undo it.

specsrbench.build.sets.training_originals(dataset)

Row indices of the un-augmented galaxies on the training side.

Parameters:

dataset (Path)

Return type:

ndarray

specsrbench.build.sets.build_eval_set(predictions, out)

eval_set.npz – the predictions plus the normalised spectra.

sigma_pix is deliberately not written. The array of that name in the historical eval_set.npz was not produced by anything in either repository and does not describe the data – roughly constant in nanometres where a spectrograph’s LSF is fixed in detector pixels, and up to 2.3x too broad at 5 um. Every classical method used it as its kernel, and with it Wiener, Tikhonov and TV merged line pairs that plain interpolation still resolves. Leaving it out means the kernel can only come from specsrbench build lsf, which measures it.

Parameters:

out (Path)

Return type:

dict

specsrbench.build.sets.main(argv=None)

Cut eval / tune / calib sets from the predictions and the paired dataset.

Return type:

int