Building the figures¶

The six paper figures are Python modules, one per figure, under specsrbench.figures. There are no notebooks.

specsrbench figures all              # all six
specsrbench figures 4                # by paper figure number
specsrbench figures mae              # or by name
specsrbench figures all --outdir /tmp/check    # somewhere other than figures/

#	name	what it shows	output
1	`toy`	every method on a 1D toy where the truth is known	`fig_toy_1d.pdf`
2	`qualitative`	one held-out galaxy, all methods, [O III] inset	`fig_jades_qualitative.pdf`
3	`residuals`	residual maps over all 572 spectra against redshift	`fig_residual_maps.pdf`
4	`mae`	global fidelity: MAE, uncertainty-normalised, RMSE	`fig_mae_summary.pdf`
5	`per-line-snr`	per-line S/N, detection, false detection, width bias	`fig_jades_per_line_snr.pdf`
6	`redshift`	error against redshift, in equal-count bins	`fig_redshift_mae.pdf`

Each prints the numbers the paper quotes from it, so a figure that is wrong is usually visible in its own output before anyone looks at the PDF.

From Python¶

from specsrbench import figures

path = figures.build("mae", outdir="/tmp")

Every figure module exposes build(cache=None, outdir=None) -> Path. Some take more: figure 2 accepts i_show to draw a different galaxy, and figure 1 takes deterministic=True and device=.

Several also expose the computation separately from the drawing, which is what the tests check:

from specsrbench.data import load_cache
from specsrbench.figures.fig5_per_line_snr import compute

rows = compute(load_cache())
rows["fdr"]["SR2"]        # false-detection rate per line

Where the data comes from¶

The figures read a cache of derived arrays — by default a cache_logR_tuned/ directory found by walking up from the working directory. Point somewhere else with:

export SPECSRBENCH_CACHE=/path/to/cache_logR_tuned

If you have no cache, Rebuilding the cache explains how to make one.

Reproducibility¶

Five of the six figures are deterministic and were verified to render pixel-identical to the published PDFs.

Figure 1 is the exception and always was. It trains a small 1D CNN inline to make its learned panel, and cuDNN reduces convolutions in a non-deterministic order, so that panel moves in the third or fourth decimal between runs — RMSE 0.0143 against 0.0144 when it was checked. No conclusion moves; the CNN wins every metric on that toy by a wide margin. Pass deterministic=True to trade some speed for a fixed result.

Warning

Rebuilding overwrites figures/*.pdf. The content is deterministic but the PDF metadata is not, so a rebuild shows up as a diff even when nothing changed. Use --outdir when you only want to check.