midas-nf-fitorientation 0.3.0

Differentiable NF-HEDM orientation and calibration fitter (drop-in replacement for FitOrientationOMP, FitOrientationParameters, FitOrientationParametersMultiPoint)

midas-nf-fitorientation

Differentiable, drop-in replacement for the three NF-HEDM orientation and calibration C executables in MIDAS:

C executable	Python equivalent	Console script
FitOrientationOMP	fit_orientation_run	midas-nf-fit-orientation
FitOrientationParameters	fit_parameters_run	midas-nf-fit-parameters
FitOrientationParametersMultiPoint	fit_multipoint_run	midas-nf-fit-multipoint

Built on top of midas-diffract, so the forward model is the same one validated to pixel-exact agreement against the C simulators. Orientation refinement uses a vectorised PyTorch Nelder-Mead running every (voxel × winner) fit problem in one batched forward call per NM iteration; calibration refinement (in fit_parameters and fit_multipoint) uses L-BFGS over the soft Gaussian-splat surrogate because Nelder-Mead scales poorly past ~10 DoF.

What changed vs. the C codes

• Orientation optimiser: NLopt Nelder-Mead → vectorised PyTorch Nelder-Mead (midas_nf_fitorientation.torch_nm). Same hard FracOverlap objective the C uses, but every (voxel, winner) fit is batched into one forward call per NM iteration. Converged simplices are trimmed out of the active set on the fly so each successive iteration is cheaper. ~22× faster than per-winner scipy NM on H100; bit-equivalent answers modulo NM convergence noise.
• Calibration optimiser (joint orientation + geometry refinement in fit_parameters / fit_multipoint): NLopt LN_NELDERMEAD → PyTorch L-BFGS over a soft Gaussian-splat surrogate with tanh-reparameterised bounds. NM is the wrong tool past ~10 DoF; L-BFGS scales well, at the cost of optimising a slightly smoothed basin floor.
• Obs volume storage: the 24 GB SpotsInfo.bin is loaded as uint8 (6 GB) by default. The orientation kernel only needs 0/1 values; the L-BFGS calibration paths request float32 explicitly.
• Bounds: each refined parameter (Eulers, tilts, Lsd, ΔLsd, BC, optionally wedge) is reparameterised as x = x0 + tol * tanh(u), so L-BFGS sees an unbounded variable but the physical parameter cannot leave its tolerance box.
• Tikhonov regularisation (opt-in, layered on top of tanh): a quadratic prior on the calibration block, useful in the multi-point joint fit where you want to drift only when many voxels' worth of evidence agree.
• Wedge refinement (opt-in, paramfile key RefineWedge 1): not available in the C code; the model accepts wedge as a calibration DoF.
• Multi-start global search for the multi-point joint fit: the C code's NM→CRS2→NM ladder is replaced with NumIterations independent L-BFGS attempts, each seeded with a Gaussian perturbation of the previous best within the tanh box. CRS2's true global behaviour is lost; for well-seeded calibration the multi-start gives the same answer in practice.
• Output files (MicFileBinary, MicFileBinary.AllMatches) match the C pwrite byte layout exactly.

Installation

From the repository root:

pip install ./packages/midas_diffract
pip install ./packages/midas_nf_fitorientation

PyTorch ≥ 2.0 is required; CUDA / MPS is auto-detected.

CLI usage

The argument signatures mirror the C executables, so existing wrapper scripts swap binaries without changes:

midas-nf-fit-orientation  params.txt blockNr nBlocks nCPUs [--device cuda] [--fp32] [--screen-only]
midas-nf-fit-parameters   params.txt rowNr   [nCPUs]      [--device cuda] [--fp32]
midas-nf-fit-multipoint   params.txt         [nCPUs]      [--device cuda] [--fp32]

Common flags (parsed after the positional args):

• --device {auto,cpu,cuda} — defaults to auto (CUDA if available).
• --fp32 — float32 forward (faster on CUDA, less bit-stable than float64 — the package default).
• --screen-only — stop after Phase 1 and dump screen_cpu.csv (mirrors MIDAS_SCREEN_ONLY=1 in the C code).
• --verbose — chatty progress.
• --lbfgs-max-iter N, --lbfgs-max-outer N — L-BFGS step limits.

Python API

import midas_nf_fitorientation as fit

# Replaces FitOrientationOMP
fit.fit_orientation_run(
    paramfile="params.txt", block_nr=0, n_blocks=1, n_cpus=8,
    device="cuda", verbose=True,
)

# Replaces FitOrientationParameters (single-voxel calibration)
result = fit.fit_parameters_run(
    paramfile="params.txt", voxel_idx=42, n_cpus=4, verbose=True,
)
print(result["Lsd"], result["tilts"], result["frac_overlap"])

# Replaces FitOrientationParametersMultiPoint (joint multi-voxel)
result = fit.fit_multipoint_run(paramfile="params.txt", n_cpus=8)

Paramfile keys

Every key consumed by the C executables is recognised, plus a small set of new keys. See midas_nf_fitorientation/params.py for the full schema. The new keys:

Key	Default	Purpose
RefineWedge 1	off	Add wedge to the calibration DoF set
WedgeTol 0.05	0.05°	Tanh-box width for wedge
TikhonovCalibration 1.0	0	Global λ for the calibration prior; 0 disables
TikhonovSigmaLsd 100.0	100 µm	Prior σ for Lsd
TikhonovSigmaTilts 0.05	0.05°	Prior σ for tilts
TikhonovSigmaBC 1.0	1 px	Prior σ for beam centres
TikhonovSigmaWedge 0.05	0.05°	Prior σ for wedge
GaussianSplatSigmaPx 1.5	auto	Override the auto-σ for the soft-overlap kernel

NumIterations (existing) controls the number of multi-start trials in fit_multipoint. Default 1 (single L-BFGS run); set to ≥ 8 for true multi-start.

Status

v0.1 is alpha. The forward path is validated against midas-diffract (pixel-exact vs. the C simulators); the fit drivers have unit-test coverage at the module level. End-to-end agreement against the C MicFileBinary on a real reconstruction dataset is the next milestone.

Licence

BSD-3-Clause.