<b>Code for "A model-fitting perspective on the robustness and speed of CLASS in reflection matrix microscopy"</b>

The three carry the two features above in different combinations, and each adjacent pair differs in one feature only, so together they form a ladder that isolates what each feature contributes.</p><p dir="ltr"><code>experiments/</code> holds the runners behind Figures 3 to 6: the multiple-scattering sweep, the aberration sweep, the reduced-angular-sampling study with its object-learning-rate sweep, and the comparison of the two CLASS update rules; <code>figures/</code> composes the four figures from their outputs.

All of them share a single protocol with fixed hyperparameters, so no solver is tuned in its own favor; the one exception is full model fit II under reduced angular sampling, whose object learning rate is chosen at each removal level as the one of eleven tested values that reaches the lowest loss, without reference to the ground truth. Wall times are measured at bare per-iteration cost, with the quality evaluation moved into a separate replay pass or kept off the solver's clock, so no reported timing carries evaluation overhead.</p><p dir="ltr"><b>Requirements and use</b>: Python 3.10 to 3.13 with NumPy, SciPy and Matplotlib, at the versions pinned in <code>requirements.txt</code> (the code itself also runs on Python 3.9).

CuPy is optional: on a GPU it returns the same results, up to floating-point rounding, at far shorter wall times. The figures and timings of the paper were produced on one NVIDIA RTX A6000 with Python 3.10.19, NumPy 2.1.2, SciPy 1.15.3 and CuPy 13.6.0 (CUDA 11.8). Only <code>config.py</code> needs editing, to point at the reflection matrices and at an output folder.

The README documents the array conventions that must be preserved, the measurement protocol, the commands for each figure, and how to call the solvers directly on a matrix of your own.</p><p dir="ltr"><b>Data</b>: This record contains code only. The 26 measured reflection matrices that the runners consume are deposited separately, at <a href="doi.org/10.6084/m9.figshare.33171494" target="_blank" rel="noreferrer">doi:10.6084/m9.figshare.33171494</a></p><p dir="ltr"><b>Version</b>: Version 2 (2.0.0, 2026-10-01) is a frozen snapshot of the code as it accompanies the revised paper; version 1 accompanied the paper as first submitted.

Changes since version 1: full model fit II (object learning rate 1e4 in the recorded units of the matrices) runs next to the other solvers in Figures 3 and 4; a new runner, <code>run_muO_sweep.py</code>, gives full model fit II of Figure 5 and the runs without the stopping rule; under reduced angular sampling the two model fits now leave the removed input modes out of the loss and out of the multiplicity count (<code>RMM_RETAINED_PAIRS=0</code> restores version 1); every runner reports the speed as the iterations and the wall time to 99 % of the recovery; <code>figures/</code> composes Figures 3 to 6; <code>requirements.txt</code> pins the versions above.

The stopping rule of the code (patience 300 iterations, relative tolerance 1e-15, cap 20000 iterations) is unchanged and is the rule of the revised paper, and outside the retained-pair treatment the solvers of version 1 compute exactly what they computed there. As in version 1, the files keep the folder structure of the repository, and they are identical to those of release v2.0.0 of the GitHub repository, <a href="github.com/Center-For-Deep-Imaging/class-model-fitting-rmm/releases/tag/v2.0.0" target="_blank" rel="noreferrer">github.com/Center-For-Deep-Imaging/class-model-fitting-rmm/releases/tag/v2.0.0</a>.

The living version, including any later corrections, is the GitHub repository: <a href="github.com/Center-For-Deep-Imaging/class-model-fitting-rmm" target="_blank" rel="noreferrer">github.com/Center-For-Deep-Imaging/class-model-fitting-rmm</a></p><p dir="ltr"><b>License</b>: Released under the MIT License</p>