onnx.numpy_helper¶

`bfloat16_to_float32`(data[, dims])	Converts ndarray of bf16 (as uint32) to f32 (as uint32).
`float8e4m3_to_float32`(data[, dims, fn, uz])	Converts ndarray of float8, e4m3 (as uint32) to f32 (as uint32).
`float8e5m2_to_float32`(data[, dims, fn, uz])	Converts ndarray of float8, e5m2 (as uint32) to f32 (as uint32).
`from_array`(tensor[, name])	Converts an array into a TensorProto including supported type defined in `onnx._custom_element_types`.
`from_dict`(dict_[, name])	Converts a Python dictionary into a map def.
`from_list`(lst[, name, dtype])	Converts a list into a sequence def.
`from_optional`(opt[, name, dtype])	Converts an optional value into a Optional def.
`to_array`(tensor[, base_dir])	Converts a tensor def object to a numpy array.
`to_dict`(map_proto)	Converts a map def to a Python dictionary.
`to_list`(sequence)	Converts a sequence def to a Python list.
`to_optional`(optional)	Converts an optional def to a Python optional.

array¶

onnx.numpy_helper.from_array(tensor: ndarray, name: str | None = None) → TensorProto[source]¶

Converts an array into a TensorProto including supported type defined in onnx._custom_element_types.

Parameters:

tensor – a numpy array.
name – (optional) the name of the tensor.

Returns:

the converted tensor def.

Return type:

TensorProto

onnx.numpy_helper.to_array(tensor: TensorProto, base_dir: str = '') → ndarray[source]¶

Converts a tensor def object to a numpy array. Supports types defined in onnx._custom_element_types.

Parameters:

tensor – a TensorProto object.
base_dir – if external tensor exists, base_dir can help to find the path to it

Returns:

the converted array.

Return type:

arr

As numpy does not support all the types defined in ONNX (float 8 types, blofat16, int4, uint4, float4e2m1), these two functions use a custom dtype defined in :mod:onnx._custom_element_types.

sequence¶

onnx.numpy_helper.to_list(sequence: SequenceProto) → list[Any][source]¶

Converts a sequence def to a Python list.

Parameters:: sequence – a SequenceProto object.
Returns:: the converted list.
Return type:: list

onnx.numpy_helper.from_list(lst: list[Any], name: str | None = None, dtype: int | None = None) → SequenceProto[source]¶

Converts a list into a sequence def.

Parameters:

lst – a Python list
name – (optional) the name of the sequence.
dtype – (optional) type of element in the input list, used for specifying sequence values when converting an empty list.

Returns:

the converted sequence def.

Return type:

SequenceProto

dictionary¶

onnx.numpy_helper.to_dict(map_proto: MapProto) → dict[Any, Any][source]¶

Converts a map def to a Python dictionary.

Parameters:: map_proto – a MapProto object.
Returns:: The converted dictionary.

onnx.numpy_helper.from_dict(dict_: dict[Any, Any], name: str | None = None) → MapProto[source]¶

Converts a Python dictionary into a map def.

Parameters:

dict – Python dictionary
name – (optional) the name of the map.

Returns:

the converted map def.

Return type:

MapProto

optional¶

onnx.numpy_helper.to_optional(optional: OptionalProto) → Any | None[source]¶

Converts an optional def to a Python optional.

Parameters:: optional – an OptionalProto object.
Returns:: the converted optional.
Return type:: opt

onnx.numpy_helper.from_optional(opt: Any | None, name: str | None = None, dtype: int | None = None) → OptionalProto[source]¶

Converts an optional value into a Optional def.

Parameters:

opt – a Python optional
name – (optional) the name of the optional.
dtype – (optional) type of element in the input, used for specifying optional values when converting empty none. dtype must be a valid OptionalProto.DataType value

Returns:

the converted optional def.

Return type:

optional

tools¶

onnx.numpy_helper.convert_endian(tensor: TensorProto) → None[source]¶

Call to convert endianness of raw data in tensor.

Parameters:: tensor – TensorProto to be converted.

onnx.numpy_helper.combine_pairs_to_complex(fa: Sequence[int]) → list[complex][source]¶

onnx.numpy_helper.create_random_int(input_shape: tuple[int], dtype: dtype, seed: int = 1) → ndarray[source]¶

Create random integer array for backend/test/case/node.

Parameters:

input_shape – The shape for the returned integer array.
dtype – The NumPy data type for the returned integer array.
seed – The seed for np.random.

Returns:

Random integer array.

Return type:

np.ndarray

onnx.numpy_helper.unpack_int4(data: int32 | ndarray, dims: int | Sequence[int], signed: bool) → ndarray[source]¶

Converts ndarray of int4 (as packed uint8) to f32 See 4 bit integer types for technical details.

Parameters:

data – A numpy array, empty dimensions are allowed if dims is None.
dims – The dimensions are used to reshape the unpacked buffer
signed – Whether the 4 bit integer is signed or unsigned

Returns:

A numpy array of float32 reshaped to dims.

cast¶

onnx.numpy_helper.bfloat16_to_float32(data: int16 | int32 | ndarray, dims: int | Sequence[int] | None = None) → ndarray[source]¶

Converts ndarray of bf16 (as uint32) to f32 (as uint32).

Parameters:

data – A numpy array, empty dimensions are allowed if dims is None.
dims – If specified, the function reshapes the results.

Returns:

A numpy array of float32 with the same dimension if dims is None, or reshaped to dims if specified

onnx.numpy_helper.float8e4m3_to_float32(data: int16 | int32 | ndarray, dims: int | Sequence[int] | None = None, fn: bool = True, uz: bool = False) → ndarray[source]¶

Converts ndarray of float8, e4m3 (as uint32) to f32 (as uint32).

See Float stored in 8 bits for technical details.

Parameters:

data – A numpy array, empty dimensions are allowed if dims is None.
dims – If specified, the function reshapes the results.
fn – No infinite values.
uz – No negative zero.

Returns:

A numpy array of float32 with the same dimension if dims is None, or reshaped to dims if specified.

onnx.numpy_helper.float8e5m2_to_float32(data: int16 | int32 | ndarray, dims: int | Sequence[int] | None = None, fn: bool = False, uz: bool = False) → ndarray[source]¶

Converts ndarray of float8, e5m2 (as uint32) to f32 (as uint32).

See Float stored in 8 bits for technical details.

Parameters:

data – A numpy array, empty dimensions are allowed if dims is None.
dims – If specified, the function reshapes the results.
fn – No infinite values.
uz – No negative zero.

Returns:

A numpy array of float32 with the same dimension if dims is None, or reshaped to dims if specified