midas-nf-preprocess 0.1.0

Differentiable PyTorch port of NF-HEDM preprocessing (CPU/CUDA/MPS): hex grid, tomo filter, diffraction-spot prediction, image processing

midas-nf-process-images

Differentiable PyTorch port of NF_HEDM/src/ProcessImagesCombined.c for CPU, CUDA, and MPS.

Pipeline

For one layer of a near-field HEDM scan:

1. Load NrFilesPerLayer raw TIFF frames into a [N, Z, Y] tensor.
2. Temporal median across frames → per-pixel background [Z, Y].
3. Per-frame:
   • Subtract median + BlanketSubtraction, clamp at 0.
   • Spatial median (3x3 or 5x5).
   • LoG response at one or more scales.
   • Threshold and label connected components → integer spot labels (detached).
   • Sigmoid surrogate → continuous spot-probability map (autograd path).
4. Accumulate spot pixels into a bit-packed SpotsInfo.bin mmap (drop-in compatible with FitOrientationOMP / simulateNF).

Phases 1, 2, and 3 (filtered, log_response, soft spot probability) are differentiable end-to-end. Connected components and the binary SpotsInfo.bin output are computed on detached tensors and do not participate in autograd.

Quickstart

from midas_nf_process_images import ProcessParams, ProcessImagesPipeline

params = ProcessParams.from_paramfile("ps.txt")
pipe = ProcessImagesPipeline(params, device="cuda")  # or "mps", "cpu"
spots = pipe.process_layer(layer_nr=1)
spots.write("/path/to/SpotsInfo.bin")

# Or all layers in one call:
spots = pipe.process_all(layers=range(1, params.n_distances + 1))

CLI

midas-nf-process-images <ParameterFile> <LayerNr> [--device cuda] [--dtype fp32] [--n-cpus 8]

Behaves like the C ProcessImagesCombined executable.

Backend selection

Same contract as midas-transforms:

• Default device: cuda -> mps -> cpu.
• Default dtype: float64 on CPU, float32 on CUDA/MPS.
• Override with device= / dtype= kwargs, or env vars MIDAS_NF_PROCESS_IMAGES_DEVICE / MIDAS_NF_PROCESS_IMAGES_DTYPE.

Differentiability

The end-to-end pipeline produces three tensors:

• filtered: median-subtracted, spatial-median-filtered image (autograd).
• log_response: LoG convolution output (autograd).
• spot_prob: soft, continuous spot-probability map via sigmoid(L / temperature) (autograd).

Plus one detached output:

• labels: integer connected-component IDs from hard-thresholded LoG (no gradient).

Optimization through the discrete spot mask uses the soft spot_prob surrogate.