Tensor

class onnx_graphsurgeon.Tensor

Bases: object

Abstract base class for tensors in a graph

This class is abstract and cannot be constructed directly.

is_empty()

Returns whether this tensor is considered empty in the graph.

Note: ‘Empty’ here refers to the name of the tensor, which is omitted for optional tensors, NOT the shape of the tensor

Returns:

Whether the tensor is empty, meaning that it is used for an omitted optional input or output.

Return type:

bool

to_constant(values: ndarray, data_location: int | None = None)

Modifies this tensor in-place to convert it to a Constant. This means that all consumers/producers of the tensor will see the update.

Parameters:

  • values (np.ndarray) – The values in this tensor

  • data_location (int) – An enum value indicating the location where the tensor data is stored. Generally, this will come from onnx.TensorProto.DataLocation.

Returns:

self

to_variable(dtype: dtype | onnx.TensorProto.DataType = None, shape: Sequence[int | str] = [])

Modifies this tensor in-place to convert it to a Variable. This means that all consumers/producers of the tensor will see the update.

Parameters:

  • dtype (Union[numpy.dtype, onnx.TensorProto.DataType]) – The data type of the tensor.

  • shape (Sequence[int]) – The shape of the tensor.

Returns:

self

i(tensor_idx=0, producer_idx=0)

Convenience function to get an input tensor of one of this tensor’s input nodes. Note that the parameters are swapped compared to the o() function; this is because tensors are likely to have only a single producer

For example:

assert tensor.i() == tensor.inputs[0].inputs[0]
assert tensor.i(1, 2) == tensor.inputs[2].inputs[1]
Parameters:

  • tensor_idx (int) – The index of the input tensor of the input node. Defaults to 0.

  • producer_idx (int) – The index of the producer node of the input tensor, if the tensor has multiple producers. Defaults to 0.

Returns:

The specified producer (input) tensor.

Return type:

Tensor

o(consumer_idx=0, tensor_idx=0)

Convenience function to get an output tensor of one of this tensor’s output nodes.

For example:

assert tensor.o() == tensor.outputs[0].outputs[0]
assert tensor.o(2, 1) == tensor.outputs[2].outputs[1]
Parameters:

  • consumer_idx (int) – The index of the consumer of the input tensor. Defaults to 0.

  • tensor_idx (int) – The index of the output tensor of the node, if the node has multiple outputs. Defaults to 0.

Returns:

The specified consumer (output) tensor

Return type:

Tensor