pyrtc.image_reconstructor

Turn wavefront-sensor images into the loop’s signal with a PyTorch model.

TorchImageReconstructor is an image-based reconstructor: it reads the wfs image stream and publishes the output of a user-supplied PyTorch model as the signal stream. It goes in the slopes section in place of SlopesProcess, so the loop, the calibration methods and the rest of the pipeline work unchanged. Typical uses are neural reconstructors and focal-plane wavefront sensing, where a network maps pixels straight to modal coefficients (run the loop with an identity interaction matrix) or to any other signal the loop calibrates as usual.

Config (slopes section):

slopes:
  class_name: TorchImageReconstructor
  signal_size: 120                 # elements of the model output
  model_file: calib/reconstructor.pt2  # torch.export (.pt2) or TorchScript
  # or a Python factory returning an nn.Module, plus optional weights:
  # model_factory: my_models:build_cnn   # "module:function"
  # model_factory_file: models.py        # then model_factory: build_cnn
  # model_kwargs: {num_outputs: 120}
  # state_dict_file: calib/weights.pt
  device: cuda:0                   # cpu (default), cuda, cuda:N
  dtype: float32                   # or float16 (CUDA only)
  cuda_graph: true                 # capture the model; eager fallback
  flux_normalization: sum          # none (default), sum or mean
  sqrt_stretch: false
  output_scale_file: ""            # .npy with signal_size factors
  functions: [compute_signal]

The work is split in two: TorchModelRunner owns the model and the real-time path (preprocessing, pinned host buffers, a dedicated CUDA stream, CUDA-graph capture) and has no streams, so it can be benchmarked and tested on its own; the component adds the stream wiring and timing.

torch is imported only when a reconstructor or runner is built, so import pyrtc stays torch-free.

Module Attributes

FLUX_EPS

Images whose normalisation flux is at or below this are treated as dark: the reconstructor publishes zeros instead of the model's response to noise.

Functions

load_torch_model(*[, model_file, ...])

Build the model a reconstructor config describes (on the CPU).

Classes

TorchImageReconstructor(conf[, model])	Publish a PyTorch model's output on each WFS image as the loop's signal.
TorchModelRunner(model, *, image_shape[, ...])	Run a PyTorch model on single WFS images, with optional CUDA graphs.