ir_data

torch_to_nnef.torch_graph.ir_data

Abstractions used in torch_to_nnef internal graph data IR.

The goal is that these elements are: - extracted/parsed from PyTorch graph data structs - translated to NNEF graph data structs

BlobTorchScriptObject dataclass

BlobTorchScriptObject(name: str, data: T.Any)

Bases: Data

Used only in Quantized Operators from our current obervation.

Data dataclass

Data(name: str, data: T.Any)

Bases: NamedItem

Base abstract T2N IR data holder.

DictTensors dataclass

DictTensors(name: str, data: T.Dict[str, TensorVariable])

Bases: Data

Used as transition object only.

None should be remaining once graph is fully expanded

FixedTensorList dataclass

FixedTensorList(name: str, data: T.Sequence[T.Union[TensorVariable, PythonConstant]])

Bases: Data

FixedTensorList is a list that contains tensor constant or not.

TensorVariable dataclass

TensorVariable(name: str, data: T.Optional[torch.Tensor], shape: T.Optional[T.List[int]], dtype: T.Optional[torch.dtype], quant: T.Optional[T.Dict[str, T.Any]] = None, _traced_data: T.Optional[torch.Tensor] = None)

Bases: Data

tracing_data property

tracing_data

Generate data if is not fixed based on tensor information.

we use it to produce computation trace

TupleTensors dataclass

TupleTensors(name: str, data: T.List[TensorVariable])

Bases: Data

Used as transition object only.

None should be remaining once graph is fully expanded