midas-ff-pipeline

End-to-end pure-Python FF-HEDM workflow. Drives all the midas-* sibling packages so a single command — or a single notebook — takes raw detector zarr inputs through to Grains.csv with no C binaries on the path.

Status: 0.2.x — dataset-density-aware GPU group_size resolver and multi-process CPU sharding (--cpu-shards) for the indexer stage; real-data shakedown on Park22 + Ti-7Al. Single-detector path is fully exercised; 1-N multi-detector pinwheel layout is supported via the existing midas_diffract.HEDMForwardModel(multi_mode="panel") backend. For the unified PF + FF orchestrator see midas-pipeline — midas-ff-pipeline and midas-pipeline run --scan-mode ff invoke the same kernels.

What it does

For each FF-HEDM layer:

The cross-detector merge is the only piece this package implements itself. It concatenates per-detector spot lists and emits a side-car Spots_det.bin mapping each obs spot to its source detector (det_id), keeping the main Spots.bin byte-compatible with the C IndexerOMP so any cross-pipeline parity test still works.

Quick start

CLI

midas-ff-pipeline run \
    --params /path/to/Parameters.txt \
    --result /path/to/run_dir \
    --layers 1-1 \
    --device cuda --dtype float64

Library API

from midas_ff_pipeline import Pipeline, PipelineConfig

pipe = Pipeline(
    config=PipelineConfig(
        result_dir="/path/to/run_dir",
        params_file="/path/to/Parameters.txt",
        n_cpus=16,
        device="cuda",
        dtype="float64",
    ),
)
pipe.run()
print(pipe.layer_result.n_grains)

Notebook

See notebooks/01_smoke_walkthrough.ipynb for the simplest end-to-end demo (synthetic 50-grain Au, single detector). 02_stage_diagnostics.ipynb adds per-stage plots between cells. 03_multi_detector_demo.ipynb runs the four-detector pinwheel layout against a synthetic ground truth.

These notebooks are not shipped with pip install — get them by cloning the MIDAS repository <https://github.com/marinerhemant/MIDAS/tree/master/packages/midas_ff_pipeline/notebooks>_.

Multi-detector

from midas_ff_pipeline import Pipeline, PipelineConfig, DetectorConfig

detectors = DetectorConfig.load_many("detectors.json")
pipe = Pipeline(
    config=PipelineConfig(
        result_dir="/path/to/multi_det_run",
        params_file="/path/to/Parameters_4det.txt",
        n_cpus=16, device="cuda", dtype="float64",
    ),
    detectors=detectors,
)
pipe.run()

detectors.json:

[
  {"det_id": 1, "zarr_path": "det1.MIDAS.zip", "lsd": 1000000.0,
   "y_bc": 1024.0, "z_bc": 1024.0, "tx": 0.0, "ty": 0.0, "tz": 0.0,
   "p_distortion": [0,0,0,0,0,0,0,0,0,0,0]},
  {"det_id": 2, "zarr_path": "det2.MIDAS.zip", "lsd": 1005000.0, ...},
  {"det_id": 3, "zarr_path": "det3.MIDAS.zip", "lsd": 1010000.0, ...},
  {"det_id": 4, "zarr_path": "det4.MIDAS.zip", "lsd": 1015000.0, ...}
]

If detectors.json is absent the pipeline falls back to DetParams 1/DetParams 2/... rows in Parameters.txt (same convention as FitMultipleGrains.c).

Resume

Each layer's directory holds a midas_state.h5 provenance ledger. --resume auto skips stages whose recorded input/output hashes match the current files. --resume from:indexing deletes the indexing-stage outputs and resumes from there. midas-ff-pipeline status prints the per-stage state for a given run dir.

Relationship to ff_MIDAS.py

FF_HEDM/workflows/ff_MIDAS.py is the historical workflow that shells out to a mix of C binaries and torch tools. midas-ff-pipeline is its pure-Python successor: same stages, no C binaries, multi-detector aware out of the box. ff_MIDAS.py is left untouched for users still on the mixed path.