Skip to content

torchsky.map.evidence

Experimental surface

These APIs are under active development. Names, defaults, and result objects may change between minor releases; pin a version if you depend on them in production.

Declared fixed-model dependencies, without tensor copies or content hashes.

Guides: likelihoods

This reference is generated from the package source. Names are grouped by the module that defines them; import every name from the public namespace torchsky.map.evidence unless the group text says otherwise.

Top-level namespace

Declared fixed-model dependencies, without tensor copies or content hashes.

EvidenceSnapshot dataclass 

EvidenceSnapshot(
    _tensors: tuple[tuple[str, Tensor, int | None], ...],
    _revision: Callable[[], object] | None = None,
    _token: object = None,
    _parents: tuple[EvidenceSnapshot, ...] = (),
    _plans: tuple[
        tuple[str, object, PlanFingerprint], ...
    ] = (),
)

A snapshot of caller-declared tensors and an optional revision getter.

Capture response matrices, precision/module parameters, calibration and support tensors explicitly. The revision getter must return an immutable scalar/tuple token and change whenever non-tensor state, tensor bindings, gauge or model conventions change. Identifiers alone do not verify physics. Dependencies retain tensors without detaching, copying or scanning them.

Torch mutation versions detect ordinary in-place writes (including shared views); NumPy and .data alias writes may bypass them. Inference tensors have no version counter and require a caller-maintained explicit revision. Arbitrary callback closures and module parameter replacement are not inspected. Rebuild factors, statistics and queries after a change.

_parents class-attribute instance-attribute 

_parents: tuple[EvidenceSnapshot, ...] = field(
    default=(), repr=False
)

_plans class-attribute instance-attribute 

_plans: tuple[tuple[str, object, PlanFingerprint], ...] = (
    field(default=(), repr=False)
)

_revision class-attribute instance-attribute 

_revision: Callable[[], object] | None = field(
    default=None, repr=False
)

_tensors class-attribute instance-attribute 

_tensors: tuple[tuple[str, Tensor, int | None], ...] = (
    field(repr=False)
)

_token class-attribute instance-attribute 

_token: object = field(default=None, repr=False)

capture classmethod 

capture(
    tensors: Mapping[str, Tensor] | None = None,
    *,
    revision: Callable[[], object] | None = None,
    plans: Mapping[str, object] | None = None,
) -> EvidenceSnapshot

validate 

validate() -> None

with_plans 

with_plans(plans: Mapping[str, object]) -> EvidenceSnapshot

Extend an unchanged snapshot with immutable graph/plan lineage.

with_tensors 

with_tensors(
    tensors: Mapping[str, Tensor],
) -> EvidenceSnapshot

Extend an unchanged snapshot, retaining its explicit revision policy.