Script Reference¶
The reproduction scripts are intentionally thin orchestration layers around the documented LION APIs. Their filenames are stable command-line entry points; output manifests record the exact expanded commands.
Pipeline¶
pipeline/PaDIS_run_pipeline.shTop-level dispatcher for GCP, Slurm, smoke, and fast-smoke modes.
pipeline/PaDIS_finalise_pipeline.shRuns unconditional generation, verification, timing extraction, tables, and figures after reconstruction.
pipeline/PaDIS_run_fast_smoke.shHardware-bounded end-to-end validation using representative models and methods without changing final settings.
Training¶
training/PaDIS_LIDC_256.pyPatch/whole-image 256-square prior training with caching, W&B resumption, timed checkpoints, and validation-intensive selection.
training/PaDIS_LIDC_512.pyNative-512 patch-prior training.
training/PaDIS_LIDC_PnP_denoiser.pyFixed-noise and noise-conditioned DRUNet training for PnP-ADMM.
Reconstruction¶
reconstruction/PaDIS_LIDC_reconstruction.pySingle-method reconstruction driver shared by matrix and tuning jobs.
reconstruction/PaDIS_run_reconstruction_matrix.pyDeclarative 109-job matrix expansion, ordering, tuned-default application, and command generation.
reconstruction/PaDIS_verify_reconstruction_matrix.pyCompleteness, settings, finite-metric, bootstrap, and optional quality-gate verification.
reconstruction/PaDIS_reconcile_reconstruction_manifest.pySemantic migration of runner markers when a manifest is rebuilt.
reconstruction/PaDIS_LIDC_generation.pySeeded unconditional prior generation.
Tuning and reporting¶
tuning/PaDIS_run_reproduction_tuning.pyandtuning/PaDIS_tune_reconstruction_hyperparameters.pyReproduce the documented validation-only candidate design.
tuning/PaDIS_hparam_defaults.pySelect and export model-safe final hyperparameter defaults.
reporting/PaDIS_make_tables.pyDecode verification metrics, extract GCP/Colab or Slurm timings, and write CSV/LaTeX tables.
reporting/PaDIS_make_paper_figures.pyRender the complete figure set with canonical labels and legacy path compatibility.
Platform launchers¶
platforms/gcp contains restartable spot-instance training and manual
multi-GPU reconstruction runners. platforms/slurm contains matching A100
array jobs, submission helpers, preflight checks, and finalisation jobs. Both
backends call the same Python training/reconstruction programs and use the same
hyperparameter registry; only scheduling, cache staging, and hardware-safe
batch limits differ.
Developer diagnostics¶
The scripts under scripts/dev check public-repository equivalence, short-
run determinism, login-node cache visibility, machine readiness, training
checkpoint resumption, and reconstruction quality. They emit structured JSON
where possible so failures can be archived alongside run logs.