mapping

Declarative weight mappings from a source file to a Keras model; importable without Keras.

SHA256constant
Transformattribute
TransposeclassPermute the axes, as numpy.transpose.
ReshapeclassReshape in C order, as numpy.reshape.
GateReorderclassReorder equal gate blocks along axis, for example ONNX LSTM "iofc" to Keras "ifco".
SplitclassTake part index of parts equal parts along axis.
SumPartsclassAdd parts equal parts along axis, for example an ONNX LSTM's input and recurrent biases.
WeightRowclassOne model weight: its source tensors, how they combine, and the transforms to its layout.
SourceFormatclassFile formats importweights reads (see heliaedge.importers.readers).
SourcePinclassThe one source file a mapping was written for.
WeightMappingclassEvery weight of one architecture, from one source format.

SHA256 = Annotated[str, Field(pattern='^[0-9a-f]{64}$')]

Transform = Annotated[Transpose | Reshape | GateReorder | Split | SumParts, Field(discriminator='kind')]

Transpose()

Permute the axes, as numpy.transpose.

model_config = ConfigDict(frozen=True, extra='forbid')

kind: Literal['transpose'] = 'transpose'

perm: tuple[int, ...]

apply(x: npt.NDArray) -> npt.NDArray

Reshape()

Reshape in C order, as numpy.reshape.

model_config = ConfigDict(frozen=True, extra='forbid')

kind: Literal['reshape'] = 'reshape'

shape: tuple[int, ...]

apply(x: npt.NDArray) -> npt.NDArray

GateReorder()

Reorder equal gate blocks along axis, for example ONNX LSTM "iofc" to Keras "ifco".

source and target name the same gates, one letter each, in their stored orders.

model_config = ConfigDict(frozen=True, extra='forbid')

kind: Literal['gate_reorder'] = 'gate_reorder'

source: str

target: str

axis: int = 0

apply(x: npt.NDArray) -> npt.NDArray

Split()

Take part index of parts equal parts along axis.

A source tensor may feed several rows only through splits that together use every part once.

model_config = ConfigDict(frozen=True, extra='forbid')

kind: Literal['split'] = 'split'

axis: int

parts: int = Field(ge=2)

index: int = Field(ge=0)

apply(x: npt.NDArray) -> npt.NDArray

SumParts()

Add parts equal parts along axis, for example an ONNX LSTM’s input and recurrent biases.

model_config = ConfigDict(frozen=True, extra='forbid')

kind: Literal['sum_parts'] = 'sum_parts'

axis: int

parts: int = Field(ge=2)

apply(x: npt.NDArray) -> npt.NDArray

WeightRow()

One model weight: its source tensors, how they combine, and the transforms to its layout.

model_config = ConfigDict(frozen=True, extra='forbid')

sources: tuple[str, ...] = Field(min_length=1)

Source tensor names. Several sources are combined by combine before transforms.

combine: Literal['sum', 'concat'] | None = None

sum (for example an LSTM’s input and recurrent biases) or concat along concat_axis; required with several sources.

concat_axis: int = 0

transforms: tuple[Transform, ...] = ()

Applied in order to the (combined) source. A Split may only come first and only with one source.

layer: str

Name of the Keras layer holding the weight (outer/inner for a nested model).

weight: str

Name of the weight within the layer, such as kernel or bias, or its path inside a composite layer, such as query/kernel in a MultiHeadAttention.

source_shape: tuple[Annotated[int, Field(ge=0)], ...] | None = None

The shape every source tensor must have, checked before combine and transforms; None checks only the final shape. A Reshape alone cannot tell a transposed source of the same size from the right one.

split: Split | None

value(tensors: dict[str, npt.NDArray]) -> npt.NDArray

This row’s weight value from the source tensors.

SourceFormat()

File formats import_weights reads (see helia_edge.importers.readers).

ONNX = 'onnx'

SAFETENSORS = 'safetensors'

TORCH = 'torch'

SourcePin()

The one source file a mapping was written for.

model_config = ConfigDict(frozen=True, extra='forbid')

uri: str

sha256: SHA256

format: SourceFormat

note: str = ''

WeightMapping()

Every weight of one architecture, from one source format.

model_config = ConfigDict(frozen=True, extra='forbid')

name: str

Mapping name.

source: SourcePin

The pinned source file and its format.

rows: tuple[WeightRow, ...] = Field(min_length=1)

One row per model weight.

unused: tuple[str, ...] = ()

Source tensors deliberately not imported (every other one must be used).