midas-transforms 0.1.0

Pure-Python/PyTorch FF-HEDM transforms (drop-in for MergeOverlappingPeaksAllZarr / CalcRadiusAllZarr / FitSetupZarr / SaveBinData)

midas-transforms

Pure-Python / PyTorch FF-HEDM intermediate transforms — the four stages between peak-fitting and indexing in the MIDAS workflow.

Drop-in replacement for these C binaries:

C binary	Python entry point
MergeOverlappingPeaksAllZarr	midas-merge-peaks
CalcRadiusAllZarr	midas-calc-radius
FitSetupZarr	midas-fit-setup
SaveBinData	midas-bin-data

Plus an end-to-end midas-transforms pipeline <zarr> that runs all four on GPU with no CSV / binary disk round-trips between stages.

Why

• Speed. Vectorised PyTorch kernels target ≥ 1× C on CPU and 5–50× on GPU at production scale. Binning and merge are the workflow's longest tails; both are now broadcast + scatter operations.
• Bit-matching with C. Every deterministic stage targets byte-exact output at float64. The geometry refine (DoFit==1) targets physics-meaningful tolerance via midas-calibrate.refine_geometry (LM with ADAM fallback — no NLopt, no Nelder-Mead).
• Differentiable. All geometry parameters that flow through detector projection (Lsd, BC_y, BC_z, tx/ty/tz, p0..p14, wedge, dLsd, dP2, residual-correction map values) carry full autograd through apply_tilt_distortion — useful for joint calibration with downstream grain mapping.
• CPU/GPU portable. Single device= switch (cpu / cuda / mps). No separate .cu codebase. No C extensions, no cibuildwheel.

Install

pip install -e packages/midas_transforms

(Until uploaded to PyPI; once released, pip install midas-transforms.)

Quick start

Per-stage CLI (drop-in for the C binaries)

midas-merge-peaks scan.zip
midas-calc-radius scan.zip
midas-fit-setup scan.zip
midas-bin-data --result-folder .

End-to-end Pipeline (intermediates stay on GPU)

midas-transforms pipeline scan.zip --device cuda --out-dir /scratch/run42

Library

from midas_transforms import Pipeline

pipe = Pipeline.from_zarr("scan.zip", device="cuda")
result = pipe.run()
result.merge        # (N, 17) on GPU
result.radius       # (N, 24) on GPU
result.fit_setup    # InputAll/Extra tensors on GPU
result.bins         # Spots/ExtraInfo/Data/nData tensors on GPU
pipe.dump("/scratch/run42")

Or use stages individually:

from midas_transforms import merge_overlapping_peaks, calc_radius, fit_setup, bin_data

merge_overlapping_peaks(zarr_path="scan.zip", device="cuda")
calc_radius(result_folder=".", zarr_params=..., device="cuda")
fit_setup(result_folder=".", zarr_params=..., device="cuda")
bin_data(result_folder=".", device="cuda")

Wiring into ff_MIDAS.py

ff_MIDAS.py accepts a --useTorchTransforms 1 flag (mirrors the existing --useTorchIndexer / --peakFitGPU flags). When set, the workflow's merge_overlaps, calc_radius, data_transform, and binning stages dispatch to the Python entry points instead of the C binaries. Outputs are byte-compatible (or numerically equivalent for the DoFit geometry refine).

Limits in v0.1.0.dev0

• Pixel-overlap merge mode (UsePixelOverlap=1) is not yet supported — use the C MergeOverlappingPeaksAllZarr for that path.
• Scanning workflows (SaveBinDataScanning, MergeMultipleScans) are out of scope here; they belong with the broader scanning pipeline.
• Byte-level regression vs C is in progress. Unit tests cover kernel correctness; full end-to-end byte-equality against C goldens lands in v0.1.0 proper.

See also

• dev/implementation_plan.md — full design, scope, and risk register.
• midas-peakfit — upstream peak-fitting (writes the consolidated HDF5).
• midas-index — downstream indexer (consumes Spots.bin/Data.bin).
• midas-calibrate — provides geometry_torch.pixel_to_REta_torch and refine_geometry (LM solver) used by fit_setup.
• midas-hkls — upstream hkls.csv generation.