midas-fit-grain

PyTorch single- and multi-grain refiner. Drop-in replacement for the C executables FitPosOrStrainsOMP / FitPosOrStrainsGPU in MIDAS FF-HEDM.

Status: 0.2.x — sparse-output pre-allocation fix (trailing-skipped seeds no longer truncate OrientPosFit / FitBest / ProcessKey) and vectorized pixel-residual fast path (~N_g× fewer kernel launches per LM iteration vs the per-grain Python loop). Park22 + datasetC CP-Ti real-data validated. The C path remains the ff_MIDAS default; this package is opt-in via --refine-backend python.

Notebooks

Worked-example Jupyter notebooks live in notebooks/. They are not shipped with pip install — get them by cloning the MIDAS repository.

What it does

For each grain in SpotsToIndex.csv:

• Reads matched spots from ExtraInfo.bin and the seed orientation from BestPos_*.csv (the indexer's per-spot output).
• Refines 12 parameters: position (3) + Bunge Euler (3) + lattice (6).
• Writes byte-identical OrientPosFit.bin / FitBest.bin / Key.bin consumed by the existing ff_MIDAS.py merge stage.

Solvers

--solver {lbfgs,adam,lm,nelder_mead} — default lbfgs.

Loss functions

--loss {pixel,angular,internal_angle} — default pixel.

Fit modes

--mode {iterative,all_at_once} — default iterative.

• iterative: position → re-match → orientation → re-match → strain → re-match → joint polish (matches FitPosOrStrainsOMP default behavior).
• all_at_once: 12 params jointly, association computed once at entry, no mid-fit re-match.

Spot weights (c-omp refiner)

Two opt-in weights on each matched spot's contribution to the stage objectives, both off by default and bit-identical to the unweighted refiner when off (verified on LSHR, bt_20id_sep26b and datasetH):

• RelFitRMSEWeightR0 r0 — scalar weight 1/(1 + rel/r0), rel = the merged spot's peak-fit FitRMSE/IMax, read from RelFitRMSE.bin (one float64 per ExtraInfo.bin row; the pipeline's refinement stage writes it when the key is set). r0 = 0.69 improves grain positions against EBSD (LSHR layer 1, ESTABLISHED 2026-09-28); it does not reduce strain noise.
• SpotWeightsDirectional 1 — experimental radial / along-ring / omega weights from SpotWeights.bin (3 float64 per row, midas_transforms.io.csv.write_spot_weights_bin). Combined with RelFitRMSEWeightR0 the orientation and position stages stay scalar and only the strain stage is directional.

A set option with a missing, short or invalid file stops the refiner with a message rather than running unweighted. Reported errors (FitBest*, DiffPos, DiffOme) stay unweighted. Measurements and preregistrations: ff-hedm LAB_NOTEBOOK.md §14.

Backends

MIDAS_FIT_GRAIN_DEVICE and MIDAS_FIT_GRAIN_DTYPE follow the same precedence contract as midas-index (cuda > mps > cpu auto-detect; f64 on CPU, f32 on accelerators). Per-grain refinement is batched into a single forward call across the block, so scaling depends on B × S (grains × spots/grain), not per-grain Python overhead.

CLI

midas-fit-grain paramstest.txt <blockNr> <numBlocks> <numLines> <numProcs> \
                [--solver lbfgs] [--loss pixel] [--mode iterative]

Argv shape mirrors the C binary so the ff_MIDAS subprocess line is one-for-one.