Does an open-source ptychography framework reproduce its own paper?¶
Reproducing: PtyRAD: A High-Performance and Flexible Ptychographic Reconstruction Framework with Automatic Differentiation (Lee et al., Microscopy and Microanalysis, 2025). DOI: 10.1093/mam/ozaf070 System: Polaris · Status: In progress
The goal¶
Check whether the authors' own open-source release of this ptychographic image-reconstruction framework actually reproduces its own published results end-to-end on independent hardware — useful both as due-diligence before adopting the framework for future imaging work, and as a template for future software-paper reproductions.
What happened¶
The framework's configuration format had changed since the paper's original snapshot, requiring a full config rewrite before anything would run. Two of six planned reconstruction targets have been rerun to completion so far, with the remaining four queued.
Results¶
- First target: per-iteration reconstruction loss trajectory matches the authors' own reference run to 3-4 significant figures.
- Second target: final reconstruction loss after 200 iterations matches the authors' bundled benchmark value to about 4 significant figures (0.3702 vs. 0.37020397).
- An informal speed comparison suggested a substantial advantage from newer GPU hardware, though this wasn't an apples-to-apples benchmark against the authors' original setup.
- Four more targets (including a regularization ablation and a hyperparameter-tuning demo) are queued to complete the full reproduction.