midas-peakfit 0.4.1

Differentiable PyTorch peak-fitting for FF-HEDM Zarr archives (drop-in replacement for PeaksFittingOMPZarrRefactor)

midas-peakfit

Differentiable PyTorch peak-fitting for FF-HEDM Zarr archives. A drop-in replacement for PeaksFittingOMPZarrRefactor (C/OpenMP/NLopt) with the following changes:

• Optimizer: Nelder-Mead → batched Levenberg-Marquardt on CPU/CUDA
• Backend: OpenMP → PyTorch (autograd, batched linear algebra)
• Precision: fp64 default; --dtype float32 available for speed
• Output: identical binary format (AllPeaks_PS.bin, AllPeaks_PX.bin)
• CLI: drop-in compatible

The C tool is kept as the validation oracle. Output is scientifically equivalent to the C tool, not bit-exact (LM and Nelder-Mead converge to slightly different minima within the same basin).

Installation

pip install -e packages/midas_peakfit[dev]

or once published:

pip install midas-peakfit

PyTorch with CUDA support must be installed separately if GPU acceleration is desired. Follow the PyTorch install guide for the right wheel for your CUDA version.

Notebooks

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

Usage

peakfit_torch DataFile.MIDAS.zip 0 1 8                    # CPU/CUDA auto
peakfit_torch DataFile.MIDAS.zip 0 1 8 --device cuda --dtype float32
peakfit_torch DataFile.MIDAS.zip 0 1 8 OutputFolder 1     # explicit ResultFolder + fitPeaks
peakfit_torch DataFile.MIDAS.zip 0 1 8 \
    --validate-against /path/to/c_AllPeaks_PS.bin         # parity check

Positional args mirror the C tool exactly: DataFile blockNr nBlocks numProcs [ResultFolder] [fitPeaks].

New flags:

Flag	Default	Meaning
--device {cpu,cuda}	cuda if available else cpu	Compute device
--dtype {float32,float64}	float64	Numeric precision
--batch-size N	4096	Cross-frame region batch threshold
--validate-against PATH	—	Compare to C-produced AllPeaks_PS.bin and emit parity report
--deterministic	off	Force deterministic algorithms (fp64 only)

Parity tolerances

Field	Tolerance
nPeaks per frame, pixel sets, maxY/maxZ, returnCode, maskTouched	exact
YCen, ZCen, Radius, diffY, diffZ	≤ 0.05 px
Eta	≤ 0.02°
IMax, IntegratedIntensity, RawSumIntensity	≤ 1% relative
BG, SigmaR, SigmaEta, σGR, σLR, σGEta, σLEta, MU, FitRMSE	≤ 5% relative

Downstream gate: indexer (IndexerOMP) on both outputs must produce identical grain orientations within 0.05° misorientation.

Output

Two binary files in {ResultFolder}/Temp/:

• AllPeaks_PS.bin — peak summary (29 columns × nPeaks per frame; see FF_HEDM/src/PeaksFittingConsolidatedIO.h for layout).
• AllPeaks_PX.bin — pixel coordinates for each peak.

These files are byte-compatible with the C tool's output and readable by ConsolidatedPeakReader/ConsolidatedPixelReader in that header.

Tests

cd packages/midas_peakfit
pytest tests/ -v                      # unit tests (fast)
pytest tests/ -v -m slow              # full pipeline parity (~3 min)
pytest tests/ -v -m gpu               # CUDA-only tests (skip on CPU)

License

BSD-3-Clause.