midas-pipeline 0.4.0

Unified MIDAS HEDM orchestrator. FF is the single-scan degeneracy of PF; one orchestrator, --scan-mode {ff,pf}.

midas-pipeline

End-to-end MIDAS HEDM orchestrator. FF is the single-scan degeneracy of PF. One package, one CLI, two scan modes.

Status

0.1.0 — end-to-end PF and FF paths live. The scanning indexer matches the C IndexerScanningOMP reference on its 1-voxel C-parity gate (seed identity, solution counts, voxel-center positions exact; orientation matrices within mrad-scale, the refiner closes the gap downstream). Real-data validation: Wenxi CP-Ti consolidation_pf reproduces the legacy pf_MIDAS.py grain count (770 == 770, all common). Park22 P5c parity gate now runs in ~6.7s vs the original 790s after the scanning-indexer position-grid fix.

Stages call in-process Python kernels via midas-index / midas-fit-grain / midas-transforms / midas-stress. FF mode shells out to python -m midas_index and python -m midas_fit_grain (same kernels, subprocess for the FF parity-preserving pattern). No CUDA C; GPU is torch-only.

Install

pip install -e packages/midas_pipeline

CLI

midas-pipeline run --scan-mode {ff,pf,auto} --params Parameters.txt --result rundir/
midas-pipeline status rundir/
midas-pipeline resume rundir/ --from <stage>
midas-pipeline reprocess rundir/
midas-pipeline inspect rundir/LayerNr_1/
midas-pipeline simulate --out simdir/ --params Parameters.txt
midas-pipeline seed --params ... --output UniqueOrientations.csv

When --scan-mode is omitted (default auto), the CLI sniffs the parameter file: nScans > 1 or presence of BeamSize / scanning keys → pf, otherwise ff. For PF mode, --n-scans, --scan-step, --beam-size, and --scan-pos-tol default to values in the params file (CLI flags override).

Indexer backend

midas-pipeline run --indexer-backend {python,c-omp} ...

python (default) — in-process numba/torch indexer. Portable (CPU/CUDA/MPS), differentiable, slower on large PF datasets.

c-omp — bundled unified C binary (midas_indexer) from midas-index. Requires midas-index installed with a working OpenMP toolchain (macOS: brew install libomp). ~290× faster than the Python path on real PF datasets (per the Wenxi CP-Ti benchmark in packages/midas_index/dev/). Output is bit-identical to the Python path on the PF parity gate.

Coexistence with midas-ff-pipeline

The legacy midas-ff-pipeline console-script is preserved as an independent FF orchestrator (its own kernels, its own CLI). It is not deprecated by midas-pipeline run --scan-mode ff — both paths invoke the same midas-index / midas-fit-grain kernels under the hood, and both stay green on the FF parity gate. Pick whichever is more convenient for your workflow.

Architecture

• One orchestrator with a mode-dependent STAGE_ORDER.
• Shared kernel packages (midas-index, midas-fit-grain, midas-transforms, etc.) extended in place; FF behavior preserved by parity gates.
• PF-only modules live inside midas_pipeline (find_grains/, sinogen, recon/, fuse, potts, em_refine, seeding/).
• Differentiability + multi-device mandatory on every new compute path (CPU / CUDA / MPS via torch).

Constraints

• No CUDA C; GPU support is torch-only.
• No deletions of legacy code in this effort.
• midas-process-grains is FF-only; PF consolidation is fresh pure-Python.
• utils/calcMiso.py is not imported by this package; all orientation math comes from midas-stress.
• TOMO/midas_tomo_python.py is imported in place, not relocated.