Types#

TensorList#

TensorList represents a batch of tensors. TensorLists are the return values of Pipeline.run(), Pipeline.outputs() or Pipeline.share_outputs().

Subsequent invocations of the mentioned functions (or Pipeline.release_outputs()) invalidate the TensorList (as well as any DALI Tensors obtained from it) and indicate to DALI that the memory can be used for something else.

TensorList wraps the outputs of current iteration and is valid only for the duration of the iteration. Using the TensorList after moving to the next iteration is not allowed. If you wish to retain the data you need to copy it before indicating DALI that you released it.

For typical use-cases, for example when DALI is used through DL Framework Plugins, no additional memory bookkeeping is necessary.

TensorListCPU#

class nvidia.dali.tensors.TensorListCPU#

__getitem__(self: TensorListCPU, i: int) → TensorCPU#: Returns a tensor at given position i in the list.

__init__(*args, **kwargs)#

Overloaded function.

__init__(self: nvidia.dali.tensors.TensorListCPU, object: capsule, layout: str = ‘’) -> None

List of tensors residing in the CPU memory.

objectDLPack object
Python DLPack object representing TensorList

layoutstr
Layout of the data
__init__(self: nvidia.dali.tensors.TensorListCPU, tl: nvidia.dali.tensors.TensorListCPU, layout: object = None) -> None
__init__(self: nvidia.dali.tensors.TensorListCPU, b: Buffer, layout: str = ‘’, is_pinned: bool = False) -> None

List of tensors residing in the CPU memory.

bobject
the buffer to wrap into the TensorListCPU object

layoutstr
Layout of the data

is_pinnedbool
If provided memory is page-locked (pinned)
__init__(self: nvidia.dali.tensors.TensorListCPU, list_of_tensors: list, layout: str = ‘’) -> None

List of tensors residing in the CPU memory.

list_of_tensors[TensorCPU]
Python list of TensorCPU objects

layoutstr
Layout of the data

as_array(self: TensorListCPU) → numpy.ndarray#: Returns TensorList as a numpy array. TensorList must be dense.

as_reshaped_tensor(self: TensorListCPU, arg0: list[int]) → TensorCPU#

Returns a tensor that is a view of this TensorList cast to the given shape.

This function can only be called if TensorList is contiguous in memory and the volumes of requested Tensor and TensorList matches.

as_tensor(self: TensorListCPU) → TensorCPU#

Returns a tensor that is a view of this TensorList.

This function can only be called if is_dense_tensor returns True.

at(self: TensorListCPU, arg0: int) → numpy.ndarray#: Returns tensor at given position in the list.

copy_to_external(self: TensorListCPU, arg0: object) → None#

Copy the contents of this TensorList to an external pointer (of type ctypes.c_void_p) residing in CPU memory.

This function is used internally by plugins to interface with tensors from supported Deep Learning frameworks.

data_ptr(self: TensorListCPU) → object#: Returns the address of the first element of TensorList.

property dtype#

Data type of the TensorListCPU’s elements.

Type:: DALIDataType

is_dense_tensor(self: TensorListCPU) → bool#

Checks whether all tensors in this TensorList have the same shape (and so the list itself can be viewed as a tensor).

For example, if TensorList contains N tensors, each with shape (H,W,C) (with the same values of H, W and C), then the list may be viewed as a tensor of shape (N, H, W, C).

layout(self: TensorListCPU) → str#

shape(self: TensorListCPU) → list[tuple]#: Shape of the tensor list.

TensorListGPU#

class nvidia.dali.tensors.TensorListGPU#

__getitem__(self: TensorListGPU, i: int) → TensorGPU#: Returns a tensor at given position i in the list.

__init__(*args, **kwargs)#

Overloaded function.

__init__(self: nvidia.dali.tensors.TensorListGPU, object: capsule, layout: str = ‘’) -> None

List of tensors residing in the GPU memory.

objectDLPack object
Python DLPack object representing TensorList

layoutstr
Layout of the data
__init__(self: nvidia.dali.tensors.TensorListGPU, tl: nvidia.dali.tensors.TensorListGPU, layout: object = None) -> None
__init__(self: nvidia.dali.tensors.TensorListGPU, list_of_tensors: list, layout: str = ‘’) -> None

List of tensors residing in the GPU memory.

list_of_tensors[TensorGPU]
Python list of TensorGPU objects

layoutstr
Layout of the data
__init__(self: nvidia.dali.tensors.TensorListGPU, object: object, layout: str = ‘’, device_id: int = -1) -> None

List of tensors residing in the GPU memory.

objectobject
Python object that implement CUDA Array Interface

layoutstr
Layout of the data

device_idint
Device of where this tensor resides. If not provided, the current device is used.
__init__(self: nvidia.dali.tensors.TensorListGPU) -> None

List of tensors residing in the GPU memory.

as_cpu(self: TensorListGPU) → TensorListCPU#: Returns a TensorListCPU object being a copy of this TensorListGPU.

as_reshaped_tensor(self: TensorListGPU, arg0: list[int]) → TensorGPU#

Returns a tensor that is a view of this TensorList cast to the given shape.

This function can only be called if TensorList is contiguous in memory and the volumes of requested Tensor and TensorList matches.

as_tensor(self: TensorListGPU) → TensorGPU#

Returns a tensor that is a view of this TensorList.

This function can only be called if is_dense_tensor returns True.

at(self: TensorListGPU, arg0: int) → TensorGPU#: Returns a tensor at given position in the list. Deprecated for __getitem__().

copy_to_external(self: TensorListGPU, ptr: object, cuda_stream: object = None, non_blocking: bool = False, use_copy_kernel: bool = False) → None#

Copy the contents of this TensorList to an external pointer residing in GPU memory.

This function is used internally by plugins to interface with tensors from supported Deep Learning frameworks.

ptrctypes.c_void_p: Destination of the copy.
cuda_streamctypes.c_void_p: CUDA stream to schedule the copy on (default stream if not provided).
non_blockingbool: Asynchronous copy.

data_ptr(self: TensorListGPU) → object#: Returns the address of the first element of TensorList.

property dtype#

Data type of the TensorListGPU’s elements.

Type:: DALIDataType

is_dense_tensor(self: TensorListGPU) → bool#

Checks whether all tensors in this TensorList have the same shape (and so the list itself can be viewed as a tensor).

For example, if TensorList contains N tensors, each with shape (H,W,C) (with the same values of H, W and C), then the list may be viewed as a tensor of shape (N, H, W, C).

layout(self: TensorListGPU) → str#

shape(self: TensorListGPU) → list[tuple]#: Shape of the tensor list.

Tensor#

TensorCPU#

class nvidia.dali.tensors.TensorCPU#

Class representing a Tensor residing in host memory. It can be used to access individual samples of a TensorListCPU or used to wrap CPU memory that is intended to be passed as an input to DALI.

It is compatible with Python Buffer Protocol and NumPy Array Interface.

dtype(self: TensorCPU) → str#: String representing NumPy type of the Tensor.

Warning

This method is deprecated. Please use TensorCPU.dtype instead.

property __array_interface__#: Returns Array Interface representation of TensorCPU.

__init__(*args, **kwargs)#

Overloaded function.

__init__(self: nvidia.dali.tensors.TensorCPU, object: capsule, layout: str = ‘’) -> None

Wrap a DLPack Tensor residing in the CPU memory.

objectDLPack object
Python DLPack object

layoutstr
Layout of the data
__init__(self: nvidia.dali.tensors.TensorCPU, b: Buffer, layout: str = ‘’, is_pinned: bool = False) -> None

Wrap a Tensor residing in the CPU memory.

bobject
the buffer to wrap into the TensorListCPU object

layoutstr
Layout of the data

is_pinnedbool
If provided memory is page-locked (pinned)

copy_to_external(self: TensorCPU, ptr: object) → None#

Copy to external pointer in the CPU memory.

ptrctypes.c_void_p: Destination of the copy.

data_ptr(self: TensorCPU) → object#: Returns the address of the first element of tensor.

property dtype#

Data type of the TensorCPU’s elements.

Type:: DALIDataType

get_property(self: TensorCPU, arg0: str) → object#

layout(self: TensorCPU) → str#

shape(self: TensorCPU) → list#: Shape of the tensor.

source_info(self: TensorCPU) → str#

Gets a string descrbing the source of the data in the tensor, e.g. a name of the file: from which the data was loaded.

squeeze(self: TensorCPU, dim: object = None) → bool#

Remove single-dimensional entries from the shape of the Tensor and it returns true if the shape changed or false if it remained unchanged.

dimint: If specified, it represents the axis of a single dimension to be squeezed.

TensorGPU#

class nvidia.dali.tensors.TensorGPU#

Class representing a Tensor residing in GPU memory. It can be used to access individual samples of a TensorListGPU or used to wrap GPU memory that is intended to be passed as an input to DALI.

It is compatible with CUDA Array Interface.

dtype(self: TensorCPU) → str#: String representing NumPy type of the Tensor.

Warning

This method is deprecated. Please use TensorGPU.dtype instead.

property __cuda_array_interface__#: Returns CUDA Array Interface (Version 2) representation of TensorGPU.

__init__(*args, **kwargs)#

Overloaded function.

__init__(self: nvidia.dali.tensors.TensorGPU, object: capsule, layout: str = ‘’) -> None

Wrap a DLPack Tensor residing in the GPU memory.

objectDLPack object
Python DLPack object

layoutstr
Layout of the data
__init__(self: nvidia.dali.tensors.TensorGPU, object: object, layout: str = ‘’, device_id: int = -1) -> None

Wrap a Tensor residing in the GPU memory that implements CUDA Array Interface.

objectobject
Python object that implements CUDA Array Interface

layoutstr
Layout of the data

device_id: int
Device of where this tensor resides. If not provided, the current device is used.

as_cpu(self: TensorGPU) → TensorCPU#: Returns a TensorCPU object being a copy of this TensorGPU.

copy_to_external(self: TensorGPU, ptr: object, cuda_stream: object = None, non_blocking: bool = False, use_copy_kernel: bool = False) → None#

Copy to external pointer in the GPU memory.

ptrctypes.c_void_p: Destination of the copy.
cuda_streamctypes.c_void_p: CUDA stream to schedule the copy on (default stream if not provided).
non_blockingbool: Asynchronous copy.

data_ptr(self: TensorGPU) → object#: Returns the address of the first element of tensor.

property dtype#

Data type of the TensorGPU’s elements.

Type:: DALIDataType

get_property(self: TensorGPU, arg0: str) → object#

layout(self: TensorGPU) → str#

shape(self: TensorGPU) → list#: Shape of the tensor.

source_info(self: TensorGPU) → str#

Gets a string descrbing the source of the data in the tensor, e.g. a name of the file: from which the data was loaded.

squeeze(self: TensorGPU, dim: object = None) → bool#

Remove single-dimensional entries from the shape of the Tensor and it returns true if the shape changed or false if it remained unchanged.

dimint: If specified, it represents the axis of a single dimension to be squeezed.

Data Layouts#

Tensor Layout String format#

DALI uses short strings (Python str type) to describe data layout in tensors, by assigning a character to each of the dimensions present in the tensor shape. For example, shape=(400, 300, 3), layout=”HWC” means that the data is an image with 3 interleaved channels, 400 pixels of height and 300 pixels of width.

For TensorLists, the index in the list is not treated as a dimension (the number of sample in the batch) and is not included in the layout.

Interpreting Tensor Layout Strings#

DALI allows you to process data of different nature (e.g. image, video, audio, volumetric images) as well as different formats (e.g. RGB image in planar configuration vs. interleaved channels). Typically, DALI operators can deal with different data formats and will behave in different way depending on the nature of the input.

While we do not restrict the valid characters to be used in a tensor layout, DALI operators assume a certain naming convention. Here is a list of commonly used dimension names:

Name	Meaning
H	Height
W	Width
C	Channels
F	Frames
D	Depth

Here are some examples of typically used layouts:

Layout	Description
HWC	Image (interleaved)
CHW	Image (planar)
DHWC	Volumetric Image (interleaved)
CDHW	Volumetric Image (planar)
FHWC	Video

For instance, a crop operation (Crop operator) receiving an input with interleaved layout (“HWC”) will infer that it should crop on the first and second dimensions (H, W). On the other hand, if the input has a planar layout (“CHW”) the crop will take place on the second and third dimensions instead.

Some operators inherently modify the layout of the data (e.g. Transpose), while others propagate the same data layout to the output (e.g. Normalize).

The layout restrictions (if any) for each operator are available through the operator’s documentation.

It is worth to note that the user is responsible to explicitly fill in the layout information when using ExternalSource API.

Constant wrapper#

Constant#

nvidia.dali.types.Constant(value, dtype=None, shape=None, layout=None, device=None, **kwargs)#

Wraps a constant value which can then be used in nvidia.dali.Pipeline.define_graph() pipeline definition step.

If the value argument is a scalar and neither shape, layout nor device is provided, the function will return a ScalarConstant wrapper object, which receives special, optimized treatment when used in Mathematical Expressions.

Otherwise, the function creates a dali.ops.Constant node, which produces a batch of constant tensors.

Parameters:

value (bool, int, float, DALIDataType DALIImageType, DALIInterpType,) – a list or tuple thereof or a numpy.ndarray The constant value to wrap. If it is a scalar, it can be used as scalar value in mathematical expressions. Otherwise, it will produce a constant tensor node (optionally reshaped according to shape argument). If this argument is is a numpy array, a PyTorch tensor or an MXNet array, the values of shape and dtype will default to value.shape and value.dtype, respectively.
dtype (DALIDataType, optional) – Target type of the constant.
shape (list or tuple of int, optional) – Requested shape of the output. If value is a scalar, it is broadcast as to fill the requested shape. Otherwise, the number of elements in value must match the volume of the shape.
layout (string, optional) – A string describing the layout of the constant tensor, e.g. “HWC”
device (string, optional, "cpu" or "gpu") – The device to place the constant tensor in. If specified, it forces the value to become a constant tensor node on given device, regardless of value type or shape.
**kwargs (additional keyword arguments) – If present, it forces the constant to become a Constant tensor node and the arguments are passed to the dali.ops.Constant operator

class nvidia.dali.types.ScalarConstant(value, dtype=None)#

Note

This class should not be instantiated directly; use Constant() function with appropriate arguments to create instances of this class.

Wrapper for a constant value that can be used in DALI Mathematical Expressions and applied element-wise to the results of DALI Operators representing Tensors in nvidia.dali.Pipeline.define_graph() step.

ScalarConstant indicates what type should the value be treated as with respect to type promotions. The actual values passed to the backend from python would be int32 for integer values and float32 for floating point values. Python builtin types bool, int and float will be marked to indicate nvidia.dali.types.DALIDataType.BOOL, nvidia.dali.types.DALIDataType.INT32, and nvidia.dali.types.DALIDataType.FLOAT respectively.

Parameters:

value (bool or int or float) – The constant value to be passed to DALI expression.
dtype (DALIDataType, optional) – Target type of the constant to be used in types promotions.

Enums#

DALIDataType#

class nvidia.dali.types.DALIDataType#

Object representing the data type of a Tensor.

NO_TYPE#

UINT8#

UINT16#

UINT32#

UINT64#

INT8#

INT16#

INT32#

INT64#

FLOAT16#

FLOAT#

FLOAT64#

BOOL#

STRING#

FEATURE#

IMAGE_TYPE#

DATA_TYPE#

INTERP_TYPE#

TENSOR_LAYOUT#

PYTHON_OBJECT#

nvidia.dali.types.to_numpy_type(dali_type)#

Converts DALIDataType to NumPy type

Parameters:: dali_type (DALIDataType) – Input type to convert

DALIIterpType#

class nvidia.dali.types.DALIInterpType#

Interpolation mode.: Note: for 2D inputs, linear and cubic are synonymous with bilinear and bicubic, respectively.

INTERP_NN#

INTERP_LINEAR#

INTERP_CUBIC#

INTERP_LANCZOS3#

INTERP_TRIANGULAR#

INTERP_GAUSSIAN#

DALIImageType#

class nvidia.dali.types.DALIImageType#

Image type

RGB#

BGR#

GRAY#

YCbCr#

ANY_DATA#

SampleInfo#

class nvidia.dali.types.SampleInfo(idx_in_epoch, idx_in_batch, iteration, epoch_idx)#

Describes the indices of a sample requested from nvidia.dali.fn.external_source()

Variables:

idx_in_epoch – 0-based index of the sample within epoch
idx_in_batch – 0-based index of the sample within batch
iteration – number of current batch within epoch
epoch_idx – number of current epoch

BatchInfo#

class nvidia.dali.types.BatchInfo(iteration, epoch_idx)#

Describes the batch requested from nvidia.dali.fn.external_source()

Variables:

iteration – number of current batch within epoch
epoch_idx – number of current epoch

TensorLayout#

class nvidia.dali.types.TensorLayout#

PipelineAPIType#

class nvidia.dali.types.PipelineAPIType(value)#

Pipeline API type

BASIC#

ITERATOR#

SCHEDULED#