torchsky.map.field.native_integrated_read¶

Conditional normalized finite-mixture law for one integrated read.

Bind with bind_native_integrated_read_likelihood. This route reuses an empty- additive FieldObservation with RAW_READS processing and the explicit single_integrated_read value semantics. Here that processing label means one continuous output per native pixel, not a nondestructive ramp, event stream, nonlinear/history response, resampled image or arbitrary pipeline residual. Fixed diagonal conversion and deterministic output offsets are declared explicitly, with their density Jacobian supplied by evaluating in observed units. CALIBRATED_OUTPUT components are allowed only in this route-local raw-read scope; legacy additive validators remain unchanged.

The native photon response maps source spectral values to photon/s in already integrated detector pixels. Its calibration input/output and quadrature output must be photon/s, with the declared physical convention photon_routing_qe_excluded. This is a caller contract for independent Poisson routing/thinning, including the fixed scale; labels cannot prove positivity/substochasticity of a callback or calibrate a sky-flat into a physical photon response. Remaining QE is an explicit [0,1] probability, with one electron at most per detected photon. Photon/charge/read mixing maps are not inferred from moments. Exposure rates are constant over the explicitly supplied positive duration; this does not integrate time history.

External typed components use RAW_READS processing. PHOTON_ILLUMINATION is [N,L] nonnegative sampled-sky spectral density, in source_value_unit with value_semantics='density_per_angstrom' and photon_shot. DARK_CHARGE is [H,W] nonnegative detector electron/s, semantics='electron_rate' and charge_shot. READ_OFFSET is [H,W] signed detector electrons, semantics='electron_offset' and deterministic_mean. CALIBRATED_OUTPUT is [H,W] signed detector output units, semantics='single_integrated_read' and deterministic_mean. Physical components are passed separately; they are not stored on FieldObservation.

Independent dark rate is added after optics, before readout. Signed read and output offsets change conditional means, never erase photon/dark shot noise. Read variance is electron^2, strictly positive on kept support. Zero read variance is a discrete law rather than this continuous density. Output calibration maps electron to observed units with no inferred gain/area/QE.

sum is the normalized independent-pixel finite-mixture NLL; mean is a reporting objective. The law, detached support and conservative first/ second derivative-domain limits are exactly those of poisson_gaussian_nll. It is not the full untruncated physical law, a Gaussian b/H/c compression, nuisance marginalization, Fisher matrix or sampling-covariance guarantee. All rates/offsets/calibration are conditional fixed evidence. Original tensors, metadata and response/calibration are guarded before/after calls; inference tensors and alias/hidden state require explicit caller revisions. Rebind after evidence changes and for a new fixed-evidence autograd graph. By default NLL evaluates all P kept rows together, with O(P*K) mixture intermediates where K can reach max_terms. Opt-in mixture_chunk_size bounds each mixture call and checkpoints its reverse-AD intermediates. Source and detector response graphs remain resident: this is not total bounded-memory survey training. Checkpointed callbacks must be deterministic and evidence unchanged through backward; the standard per-call evidence guards apply.

_snapshot: EvidenceSnapshot = field(
    init=False, repr=False, compare=False
)

_support: Tensor = field(
    init=False, repr=False, compare=False
)

additive_components: tuple[AdditiveComponent, ...] = ()

calibration_id: str

dependencies: EvidenceSnapshot | None = None

electrons_per_photon: Tensor

exposure_time_s: Tensor

max_terms: int = 4096

measurement_ids: tuple[str, ...]

observation: FieldObservation

output_calibration: PixelCalibrationOp

photon_response: NativeFieldResponse

provenance_id: str

read_variance_electrons2: Tensor

revision: Callable[[], object] | None = None

tail_sigma: float = 12.0

__post_init__() -> None

_law_parameters(
    values: Tensor,
) -> tuple[Tensor, Tensor, Tensor, Tensor]

_validate_component(component: AdditiveComponent) -> None

expected_output(values: Tensor) -> Tensor

nll(
    values: Tensor,
    *,
    reduction: Literal["sum", "mean", "none"] = "sum",
    mixture_chunk_size: int | None = None,
) -> Tensor

nll_field(
    model: Module,
    *,
    forward_context: ForwardContext | None = None,
    source_chunk_size: int = 4096,
    reduction: Literal["sum", "mean", "none"] = "sum",
    mixture_chunk_size: int | None = None,
) -> Tensor

validate_evidence() -> None