Tensors

IOBinding

class onnxruntime.IOBinding(session)

This class provides API to bind input/output to a specified device, e.g. GPU.

bind_cpu_input(name, arr_on_cpu)

bind an input to array on CPU :param name: input name :param arr_on_cpu: input values as a python array on CPU

bind_input(name, device_type, device_id, element_type, shape, buffer_ptr)

Parameters:

• name – input name

• device_type – e.g. cpu, cuda

• device_id – device id, e.g. 0

• element_type – input element type

• shape – input shape

• buffer_ptr – memory pointer to input data

bind_ortvalue_input(name, ortvalue)

Parameters:

• name – input name

• ortvalue – OrtValue instance to bind

bind_ortvalue_output(name, ortvalue)

Parameters:

• name – output name

• ortvalue – OrtValue instance to bind

bind_output(name, device_type='cpu', device_id=0, element_type=None, shape=None, buffer_ptr=None)

Parameters:

• name – output name

• device_type – e.g. cpu, cuda, cpu by default

• device_id – device id, e.g. 0

• element_type – output element type

• shape – output shape

• buffer_ptr – memory pointer to output data

copy_outputs_to_cpu()

Copy output contents to CPU (if on another device). No-op if already on the CPU.

get_outputs()

Returns the output OrtValues from the Run() that preceded the call. The data buffer of the obtained OrtValues may not reside on CPU memory

OrtDevice

class onnxruntime.OrtDevice(c_ort_device)

A data structure that exposes the underlying C++ OrtDevice

Internal constructor

OrtValue

class onnxruntime.OrtValue(ortvalue, numpy_obj=None)

A data structure that supports all ONNX data formats (tensors and non-tensors) that allows users to place the data backing these on a device, for example, on a CUDA supported device. This class provides APIs to construct and deal with OrtValues.

as_sparse_tensor()

The function will return SparseTensor contained in this OrtValue

data_ptr()

Returns the address of the first element in the OrtValue’s data buffer

data_type()

Returns the data type of the data in the OrtValue

device_name()

Returns the name of the device where the OrtValue’s data buffer resides e.g. cpu, cuda

element_type()

Returns the proto type of the data in the OrtValue if the OrtValue is a tensor.

has_value()

Returns True if the OrtValue corresponding to an optional type contains data, else returns False

is_sparse_tensor()

Returns True if the OrtValue contains a SparseTensor, else returns False

is_tensor()

Returns True if the OrtValue contains a Tensor, else returns False

is_tensor_sequence()

Returns True if the OrtValue contains a Tensor Sequence, else returns False

numpy()

Returns a Numpy object from the OrtValue. Valid only for OrtValues holding Tensors. Throws for OrtValues holding non-Tensors. Use accessors to gain a reference to non-Tensor objects such as SparseTensor

static ort_value_from_sparse_tensor(sparse_tensor)

The function will construct an OrtValue instance from a valid SparseTensor The new instance of OrtValue will assume the ownership of sparse_tensor

static ortvalue_from_numpy(numpy_obj, device_type='cpu', device_id=0)

Factory method to construct an OrtValue (which holds a Tensor) from a given Numpy object A copy of the data in the Numpy object is held by the OrtValue only if the device is NOT cpu

Parameters:

• numpy_obj – The Numpy object to construct the OrtValue from

• device_type – e.g. cpu, cuda, cpu by default

• device_id – device id, e.g. 0

static ortvalue_from_shape_and_type(shape=None, element_type=None, device_type='cpu', device_id=0)

Factory method to construct an OrtValue (which holds a Tensor) from given shape and element_type

Parameters:

• shape – List of integers indicating the shape of the OrtValue

• element_type – The data type of the elements in the OrtValue (numpy type)

• device_type – e.g. cpu, cuda, cpu by default

• device_id – device id, e.g. 0

shape()

Returns the shape of the data in the OrtValue

update_inplace(np_arr)

Update the OrtValue in place with a new Numpy array. The numpy contents are copied over to the device memory backing the OrtValue. It can be used to update the input valuess for an InferenceSession with CUDA graph enabled or other scenarios where the OrtValue needs to be updated while the memory address can not be changed.

SparseTensor

class onnxruntime.SparseTensor(sparse_tensor)

A data structure that project the C++ SparseTensor object The class provides API to work with the object. Depending on the format, the class will hold more than one buffer depending on the format

Internal constructor

as_blocksparse_view()

The method will return coo representation of the sparse tensor which will enable querying BlockSparse indices. If the instance did not contain BlockSparse format, it would throw. You can query coo indices as:

block_sparse_indices = sparse_tensor.as_blocksparse_view().indices()

which will return a numpy array that is backed by the native memory

as_coo_view()

The method will return coo representation of the sparse tensor which will enable querying COO indices. If the instance did not contain COO format, it would throw. You can query coo indices as:

coo_indices = sparse_tensor.as_coo_view().indices()

which will return a numpy array that is backed by the native memory.

as_csrc_view()

The method will return CSR(C) representation of the sparse tensor which will enable querying CRS(C) indices. If the instance dit not contain CSR(C) format, it would throw. You can query indices as:

inner_ndices = sparse_tensor.as_csrc_view().inner()
outer_ndices = sparse_tensor.as_csrc_view().outer()

returning numpy arrays backed by the native memory.

data_type()

Returns a string data type of the data in the OrtValue

dense_shape()

Returns a numpy array(int64) containing a dense shape of a sparse tensor

device_name()

Returns the name of the device where the SparseTensor data buffers reside e.g. cpu, cuda

format()

Returns a OrtSparseFormat enumeration

static sparse_coo_from_numpy(dense_shape, values, coo_indices, ort_device)

Factory method to construct a SparseTensor in COO format from given arguments

Parameters:

• dense_shape – 1-D numpy array(int64) or a python list that contains a dense_shape of the sparse tensor must be on cpu memory

• values – a homogeneous, contiguous 1-D numpy array that contains non-zero elements of the tensor of a type.

• coo_indices – contiguous numpy array(int64) that contains COO indices for the tensor. coo_indices may have a 1-D shape when it contains a linear index of non-zero values and its length must be equal to that of the values. It can also be of 2-D shape, in which has it contains pairs of coordinates for each of the nnz values and its length must be exactly twice of the values length.

• ort_device –

  • describes the backing memory owned by the supplied nummpy arrays. Only CPU memory is

  suppored for non-numeric data types.

For primitive types, the method will map values and coo_indices arrays into native memory and will use them as backing storage. It will increment the reference count for numpy arrays and it will decrement it on GC. The buffers may reside in any storage either CPU or GPU. For strings and objects, it will create a copy of the arrays in CPU memory as ORT does not support those on other devices and their memory can not be mapped.

static sparse_csr_from_numpy(dense_shape, values, inner_indices, outer_indices, ort_device)

Factory method to construct a SparseTensor in CSR format from given arguments

Parameters:

• dense_shape – 1-D numpy array(int64) or a python list that contains a dense_shape of the sparse tensor (rows, cols) must be on cpu memory

• values – a contiguous, homogeneous 1-D numpy array that contains non-zero elements of the tensor of a type.

• inner_indices – contiguous 1-D numpy array(int64) that contains CSR inner indices for the tensor. Its length must be equal to that of the values.

• outer_indices – contiguous 1-D numpy array(int64) that contains CSR outer indices for the tensor. Its length must be equal to the number of rows + 1.

• ort_device –

  • describes the backing memory owned by the supplied nummpy arrays. Only CPU memory is

  suppored for non-numeric data types.

For primitive types, the method will map values and indices arrays into native memory and will use them as backing storage. It will increment the reference count and it will decrement then count when it is GCed. The buffers may reside in any storage either CPU or GPU. For strings and objects, it will create a copy of the arrays in CPU memory as ORT does not support those on other devices and their memory can not be mapped.

to_cuda(ort_device)

Returns a copy of this instance on the specified cuda device

Parameters:

ort_device – with name ‘cuda’ and valid gpu device id

The method will throw if:

• this instance contains strings

• this instance is already on GPU. Cross GPU copy is not supported

• CUDA is not present in this build

• if the specified device is not valid

values()

The method returns a numpy array that is backed by the native memory if the data type is numeric. Otherwise, the returned numpy array that contains copies of the strings.