Operation Reference¶

The data processing graph within a DALI Pipeline is defined by calling operation functions. They accept and return instances of DataNode, which are symbolic representations of batches of Tensors. The operation functions cannot be used to process data directly.

The constraints for defining the processing pipeline can be found in this section of Pipeline documentation.

The following table lists all available operations available in DALI:

Function	Device support	Short description
`audio_decoder`	CPU	Legacy alias for `decoders.audio()`.
`batch_permutation`	CPU	Produces a batch of random integers which can be used as indices for indexing samples in the batch.
`bb_flip`	CPU, GPU	Flips bounding boxes horizontaly or verticaly (mirror).
`bbox_paste`	CPU	Transforms bounding boxes so that the boxes remain in the same place in the image after the image is pasted on a larger canvas.
`box_encoder`	CPU, GPU	Encodes the input bounding boxes and labels using a set of default boxes (anchors) passed as an argument.
`brightness`	CPU, GPU	Adjusts the brightness of the images.
`brightness_contrast`	CPU, GPU	Adjusts the brightness and contrast of the images.
`caffe2_reader`	CPU	Legacy alias for `readers.caffe2()`.
`caffe_reader`	CPU	Legacy alias for `readers.caffe()`.
`cast`	CPU, GPU	Cast tensor to a different type.
`cat`	CPU, GPU	Joins the input tensors along an existing axis.
`coco_reader`	CPU	Legacy alias for `readers.coco()`.
`coin_flip`	CPU, GPU	Generates random boolean values following a bernoulli distribution.
`color_space_conversion`	CPU, GPU	Converts between various image color models.
`color_twist`	CPU, GPU	Adjusts hue, saturation, brightness and contrast of the image.
`compose`	CPU, GPU	Returns a meta-operator that chains the operations in op_list.
`constant`	CPU, GPU	Produces a batch of constant tensors.
`contrast`	CPU, GPU	Adjusts the contrast of the images.
`coord_flip`	CPU, GPU	Transforms vectors or points by flipping (reflecting) their coordinates with respect to a given center.
`coord_transform`	CPU, GPU	Applies a linear transformation to points or vectors.
`copy`	CPU, GPU	Creates a copy of the input tensor.
`crop`	CPU, GPU	Crops the images with the specified window dimensions and window position (upper left corner).
`crop_mirror_normalize`	CPU, GPU	Performs fused cropping, normalization, format conversion (NHWC to NCHW) if desired, and type casting.
`dl_tensor_python_function`	CPU, GPU	Executes a Python function that operates on DLPack tensors.
`dump_image`	CPU, GPU	Save images in batch to disk in PPM format.
`element_extract`	CPU, GPU	Extracts one or more elements from input sequence.
`erase`	CPU, GPU	Erases one or more regions from the input tensors.
`expand_dims`	CPU, GPU	Insert new dimension(s) with extent 1 to the data shape.
`external_source`	CPU, GPU	Allows externally provided data to be passed as an input to the pipeline.
`fast_resize_crop_mirror`	CPU	Performs a fused resize, crop, mirror operation.
`file_reader`	CPU	Legacy alias for `readers.file()`.
`flip`	CPU, GPU	Flips the images in selected dimensions (horizontal, vertical, and depthwise).
`gaussian_blur`	CPU, GPU	Applies a Gaussian Blur to the input.
`get_property`	CPU, GPU	Returns a property of the tensor passed as an input.
`grid_mask`	CPU, GPU	Performs the gridmask augumentation (https://arxiv.org/abs/2001.04086).
`hsv`	CPU, GPU	Adjusts hue, saturation and value (brightness) of the images.
`hue`	CPU, GPU	Changes the hue level of the image.
`image_decoder`	CPU, Mixed	Legacy alias for `decoders.image()`.
`image_decoder_crop`	CPU, Mixed	Legacy alias for `decoders.image_crop()`.
`image_decoder_random_crop`	CPU, Mixed	Legacy alias for `decoders.image_random_crop()`.
`image_decoder_slice`	CPU, Mixed	Legacy alias for `decoders.image_slice()`.
`jitter`	GPU	Performs a random Jitter augmentation.
`jpeg_compression_distortion`	CPU, GPU	Introduces JPEG compression artifacts to RGB images.
`laplacian`	CPU, GPU	Computes the Laplacian of an input.
`lookup_table`	CPU, GPU	Maps the input to output by using a lookup table that is specified by `keys` and `values`, and a `default_value` for unspecified keys.
`mel_filter_bank`	CPU, GPU	Converts a spectrogram to a mel spectrogram by applying a bank of triangular filters.
`mfcc`	CPU, GPU	Computes Mel Frequency Cepstral Coefficiencs (MFCC) from a mel spectrogram.
`multi_paste`	CPU, GPU	Performs multiple pastes from image batch to each of outputs
`mxnet_reader`	CPU	Legacy alias for `readers.mxnet()`.
`nemo_asr_reader`	CPU	Legacy alias for `readers.nemo_asr()`.
`nonsilent_region`	CPU	Performs leading and trailing silence detection in an audio buffer.
`normal_distribution`	CPU, GPU	Generates random numbers following a normal distribution.
`normalize`	CPU, GPU	Normalizes the input by removing the mean and dividing by the standard deviation.
`numba_function`	CPU	Invokes a njit compiled Numba function.
`numpy_reader`	CPU, GPU	Legacy alias for `readers.numpy()`.
`one_hot`	CPU, GPU	Produces a one-hot encoding of the input.
`optical_flow`	GPU	Calculates the optical flow between images in the input.
`pad`	CPU, GPU	Pads all samples with the `fill_value` in the specified axes to match the biggest extent in the batch for those axes or to match the minimum shape specified.
`paste`	GPU	Pastes the input images on a larger canvas, where the canvas size is equal to `input size * ratio`.
`peek_image_shape`	CPU	Obtains the shape of the encoded image.
`permute_batch`	CPU, GPU	Returns a batch of tensors constructed by selecting tensors from the input based on indices given in `indices` argument:
`power_spectrum`	CPU	Calculates power spectrum of the signal.
`preemphasis_filter`	CPU, GPU	Applies preemphasis filter to the input data.
`python_function`	CPU, GPU	Executes a Python function.
`random_bbox_crop`	CPU	Applies a prospective random crop to an image coordinate space while keeping the bounding boxes, and optionally labels, consistent.
`random_resized_crop`	CPU, GPU	Performs a crop with a randomly selected area and aspect ratio and resizes it to the specified size.
`reinterpret`	CPU, GPU	Treats content of the input as if it had a different type, shape, and/or layout.
`reshape`	CPU, GPU	Treats content of the input as if it had a different shape and/or layout.
`resize`	CPU, GPU	Resize images.
`resize_crop_mirror`	CPU	Performs a fused resize, crop, mirror operation. Both fixed and random resizing and cropping are supported.
`roi_random_crop`	CPU	Produces a fixed shape cropping window, randomly placed so that as much of the provided region of interest (ROI) is contained in it.
`rotate`	CPU, GPU	Rotates the images by the specified angle.
`saturation`	CPU, GPU	Changes the saturation level of the image.
`sequence_reader`	CPU	Legacy alias for `readers.sequence()`.
`sequence_rearrange`	CPU, GPU	Rearranges frames in a sequence.
`shapes`	CPU, GPU	Returns the shapes of inputs.
`slice`	CPU, GPU	Extracts a subtensor, or slice.
`spectrogram`	CPU, GPU	Produces a spectrogram from a 1D signal (for example, audio).
`sphere`	CPU, GPU	Performs a sphere augmentation.
`squeeze`	CPU, GPU	Removes the dimensions given as `axes` or `axis_names`.
`ssd_random_crop`	CPU	Performs a random crop with bounding boxes where Intersection Over Union (IoU) meets a randomly selected threshold between 0-1.
`stack`	CPU, GPU	Joins the input tensors along a new axis.
`tfrecord_reader`	CPU	Legacy alias for `readers.tfrecord()`.
`to_decibels`	CPU, GPU	Converts a magnitude (real, positive) to the decibel scale by using the following formula:
`torch_python_function`	CPU, GPU	Executes a function that is operating on Torch tensors.
`transpose`	CPU, GPU	Transposes the tensors by reordering the dimensions based on the `perm` parameter.
`uniform`	CPU, GPU	Generates random numbers following a uniform distribution.
`video_reader`	GPU	Legacy alias for `readers.video()`.
`video_reader_resize`	GPU	Legacy alias for `readers.video_resize()`.
`warp_affine`	CPU, GPU	Applies an affine transformation to the images.
`water`	CPU, GPU	Performs a water augmentation, which makes the image appear to be underwater.
`decoders.audio`	CPU	Decodes waveforms from encoded audio data.
`decoders.image`	CPU, Mixed	Decodes images.
`decoders.image_crop`	CPU, Mixed	Decodes images and extracts regions-of-interest (ROI) that are specified by fixed window dimensions and variable anchors.
`decoders.image_random_crop`	CPU, Mixed	Decodes images and randomly crops them.
`decoders.image_slice`	CPU, Mixed	Decodes images and extracts regions of interest.
`experimental.readers.video`	CPU	Loads and decodes video files using FFmpeg.
`noise.gaussian`	CPU, GPU	Applies gaussian noise to the input.
`noise.salt_and_pepper`	CPU, GPU	Applies salt-and-pepper noise to the input.
`noise.shot`	CPU, GPU	Applies shot noise to the input.
`random.coin_flip`	CPU, GPU	Generates random boolean values following a bernoulli distribution.
`random.normal`	CPU, GPU	Generates random numbers following a normal distribution.
`random.uniform`	CPU, GPU	Generates random numbers following a uniform distribution.
`readers.caffe`	CPU	Reads (Image, label) pairs from a Caffe LMDB.
`readers.caffe2`	CPU	Reads sample data from a Caffe2 Lightning Memory-Mapped Database (LMDB).
`readers.coco`	CPU	Reads data from a COCO dataset that is composed of a directory with images and annotation JSON files.
`readers.file`	CPU	Reads file contents and returns file-label pairs.
`readers.mxnet`	CPU	Reads the data from an MXNet RecordIO.
`readers.nemo_asr`	CPU	Reads automatic speech recognition (ASR) data (audio, text) from an NVIDIA NeMo compatible manifest.
`readers.numpy`	CPU, GPU	Reads Numpy arrays from a directory.
`readers.sequence`	CPU	Reads [Frame] sequences from a directory representing a collection of streams.
`readers.tfrecord`	CPU	Reads samples from a TensorFlow TFRecord file.
`readers.video`	GPU	Loads and decodes video files using FFmpeg and NVDECODE, which is the hardware-accelerated video decoding feature in the NVIDIA(R) GPU.
`readers.video_resize`	GPU	Loads, decodes and resizes video files with FFmpeg and NVDECODE, which is NVIDIA GPU’s hardware-accelerated video decoding.
`readers.webdataset`	CPU	A reader for the webdataset format.
`reductions.max`	CPU, GPU	Gets maximal input element along provided axes.
`reductions.mean`	CPU, GPU	Gets mean of elements along provided axes.
`reductions.mean_square`	CPU, GPU	Gets mean square of elements along provided axes.
`reductions.min`	CPU, GPU	Gets minimal input element along provided axes.
`reductions.rms`	CPU, GPU	Gets root mean square of elements along provided axes.
`reductions.std_dev`	CPU, GPU	Gets standard deviation of elements along provided axes.
`reductions.sum`	CPU, GPU	Gets sum of elements along provided axes.
`reductions.variance`	CPU, GPU	Gets variance of elements along provided axes.
`segmentation.random_mask_pixel`	CPU	Selects random pixel coordinates in a mask, sampled from a uniform distribution.
`segmentation.random_object_bbox`	CPU	Randomly selects an object from a mask and returns its bounding box.
`segmentation.select_masks`	CPU	Selects a subset of polygons by their mask ids.
`transforms.combine`	CPU	Combines two or more affine transforms.
`transforms.crop`	CPU	Produces an affine transform matrix that maps a reference coordinate space to another one.
`transforms.rotation`	CPU	Produces a rotation affine transform matrix.
`transforms.scale`	CPU	Produces a scale affine transform matrix.
`transforms.shear`	CPU	Produces a shear affine transform matrix.
`transforms.translation`	CPU	Produces a translation affine transform matrix.

nvidia.dali.fn¶

nvidia.dali.fn.audio_decoder(*inputs, **kwargs)¶: Warning

This operator is now deprecated. Use decoders.audio() instead.

In DALI 1.0 all decoders were moved into a dedicated decoders submodule and renamed to follow a common pattern. This is a placeholder operator with identical functionality to allow for backward compatibility.

Legacy alias for decoders.audio().

nvidia.dali.fn.batch_permutation(*inputs, **kwargs)¶

Produces a batch of random integers which can be used as indices for indexing samples in the batch.

Supported backends

‘cpu’

Keyword Arguments

allow_repetitions (bool, optional, default = False) – If true, the output can contain repetitions and omissions.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
no_fixed_points (bool, optional, default = False) – If true, the the output permutation cannot contain fixed points, that is out[i] != i. This argument is ignored when batch size is 1.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.bb_flip(*inputs, **kwargs)¶

Flips bounding boxes horizontaly or verticaly (mirror).

The bounding box coordinates for the input are in the [x, y, width, height] - xywh or [left, top, right, bottom] - ltrb format. All coordinates are in the image coordinate system, that is 0.0-1.0

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
horizontal (int or TensorList of int, optional, default = 1) – Flip horizontal dimension.
ltrb (bool, optional, default = False) – True for ltrb or False for xywh.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
vertical (int or TensorList of int, optional, default = 0) – Flip vertical dimension.

nvidia.dali.fn.bbox_paste(*inputs, **kwargs)¶

Transforms bounding boxes so that the boxes remain in the same place in the image after the image is pasted on a larger canvas.

Corner coordinates are transformed according to the following formula:

(x', y') = (x/ratio + paste_x', y/ratio + paste_y')

Box sizes (if xywh is used) are transformed according to the following formula:

(w', h') = (w/ratio, h/ratio)

Where:

paste_x' = paste_x * (ratio - 1)/ratio
paste_y' = paste_y * (ratio - 1)/ratio

The paste coordinates are normalized so that (0,0) aligns the image to top-left of the canvas and (1,1) aligns it to bottom-right.

Supported backends

‘cpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

ratio (float or TensorList of float) – Ratio of the canvas size to the input size; the value must be at least 1.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
ltrb (bool, optional, default = False) – True for ltrb or False for xywh.
paste_x (float or TensorList of float, optional, default = 0.5) – Horizontal position of the paste in image coordinates (0.0 - 1.0).
paste_y (float or TensorList of float, optional, default = 0.5) – Vertical position of the paste in image coordinates (0.0 - 1.0).
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.box_encoder(*inputs, **kwargs)¶

Encodes the input bounding boxes and labels using a set of default boxes (anchors) passed as an argument.

This operator follows the algorithm described in “SSD: Single Shot MultiBox Detector” and implemented in https://github.com/mlperf/training/tree/master/single_stage_detector/ssd. Inputs must be supplied as the following Tensors:

BBoxes that contain bounding boxes that are represented as [l,t,r,b].
Labels that contain the corresponding label for each bounding box.

The results are two tensors:

EncodedBBoxes that contain M-encoded bounding boxes as [l,t,r,b], where M is number of anchors.
EncodedLabels that contain the corresponding label for each encoded box.

Supported backends

‘cpu’
‘gpu’

Parameters

input0 (TensorList) – Input to the operator.
input1 (TensorList) – Input to the operator.

Keyword Arguments

anchors (float or list of float) – Anchors to be used for encoding, as the list of floats is in the ltrb format.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
criteria (float, optional, default = 0.5) –
Threshold IoU for matching bounding boxes with anchors.

The value needs to be between 0 and 1.
means (float or list of float, optional, default = [0.0, 0.0, 0.0, 0.0]) – [x y w h] mean values for normalization.
offset (bool, optional, default = False) – Returns normalized offsets ((encoded_bboxes*scale - anchors*scale) - mean) / stds in EncodedBBoxes that use std and the mean and scale arguments.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
scale (float, optional, default = 1.0) – Rescales the box and anchor values before the offset is calculated (for example, to return to the absolute values).
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
stds (float or list of float, optional, default = [1.0, 1.0, 1.0, 1.0]) – [x y w h] standard deviations for offset normalization.

nvidia.dali.fn.brightness(*inputs, **kwargs)¶

Adjusts the brightness of the images.

The brightness is adjusted based on the following formula:

out = brightness_shift * output_range + brightness * in

Where output_range is 1 for float outputs or the maximum positive value for integral types.

This operator can also change the type of data.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList ('FHWC', 'DHWC', 'HWC')) – Input to the operator.

Keyword Arguments

brightness (float or TensorList of float, optional, default = 1.0) – Brightness mutliplier.
brightness_shift (float or TensorList of float, optional, default = 0.0) –
The brightness shift.

For signed types, 1.0 represents the maximum positive value that can be represented by the type.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.NO_TYPE) –
Output data type.

If not set, the input type is used.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.brightness_contrast(*inputs, **kwargs)¶

Adjusts the brightness and contrast of the images.

The brightness and contrast are adjusted based on the following formula:

out = brightness_shift * output_range + brightness * (contrast_center + contrast * (in - contrast_center))

Where the output_range is 1 for float outputs or the maximum positive value for integral types.

This operator can also change the type of data.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList ('FHWC', 'DHWC', 'HWC')) – Input to the operator.

Keyword Arguments

brightness (float or TensorList of float, optional, default = 1.0) – Brightness mutliplier.
brightness_shift (float or TensorList of float, optional, default = 0.0) –
The brightness shift.

For signed types, 1.0 represents the maximum positive value that can be represented by the type.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
contrast (float or TensorList of float, optional, default = 1.0) – The contrast multiplier, where 0.0 produces the uniform grey.
contrast_center (float, optional, default = 0.5) –
The intensity level that is unaffected by contrast.

This is the value that all pixels assume when the contrast is zero. When not set, the half of the input type’s positive range (or 0.5 for float) is used.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.NO_TYPE) –
Output data type.

If not set, the input type is used.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.caffe2_reader(*inputs, **kwargs)¶: Warning

This operator is now deprecated. Use readers.caffe2() instead.

In DALI 1.0 all readers were moved into a dedicated readers submodule and renamed to follow a common pattern. This is a placeholder operator with identical functionality to allow for backward compatibility.

Legacy alias for readers.caffe2().

nvidia.dali.fn.caffe_reader(*inputs, **kwargs)¶: Warning

This operator is now deprecated. Use readers.caffe() instead.

In DALI 1.0 all readers were moved into a dedicated readers submodule and renamed to follow a common pattern. This is a placeholder operator with identical functionality to allow for backward compatibility.

Legacy alias for readers.caffe().

nvidia.dali.fn.cast(*inputs, **kwargs)¶

Cast tensor to a different type.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

dtype (nvidia.dali.types.DALIDataType) – Output data type.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.cat(*inputs, **kwargs)¶

Joins the input tensors along an existing axis.

The shapes of the inputs must match in all dimensions except the concatenation axis.

Supported backends

‘cpu’
‘gpu’

Parameters

input0 (TensorList) – Input to the operator.
input[1..998] (TensorList, optional) – This function accepts up to 998 optional positional inputs

Keyword Arguments

axis (int, optional, default = 0) – Axis along which the input tensors are concatenated.
axis_name (str, optional) –
Name of the axis along which the tensors are concatenated.

This argument is mutually exclusive with axis. This argument requires that at least one input has a non-empty layout and that all non-empty input layouts match.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.coco_reader(*inputs, **kwargs)¶: Warning

This operator is now deprecated. Use readers.coco() instead.

In DALI 1.0 all readers were moved into a dedicated readers submodule and renamed to follow a common pattern. This is a placeholder operator with identical functionality to allow for backward compatibility.

Legacy alias for readers.coco().

nvidia.dali.fn.coin_flip(*inputs, **kwargs)¶

Warning

This operator is now deprecated. Use random.coin_flip() instead.

Generates random boolean values following a bernoulli distribution.

The probability of generating a value 1 (true) is determined by the probability argument.

The shape of the generated data can be either specified explicitly with a shape argument, or chosen to match the shape of the input, if provided. If none are present, a single value per sample is generated.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList, optional) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional) –
Output data type.

Note

The generated numbers are converted to the output data type, rounding and clamping if necessary.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
probability (float or TensorList of float, optional, default = 0.5) – Probability of value 1.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shape (int or list of int or TensorList of int, optional) – Shape of the output data.

nvidia.dali.fn.color_space_conversion(*inputs, **kwargs)¶

Converts between various image color models.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList ('FDHWC', 'FHWC', 'DHWC', 'HWC')) – Input to the operator.

Keyword Arguments

image_type (nvidia.dali.types.DALIImageType) – The color space of the input image.
output_type (nvidia.dali.types.DALIImageType) – The color space of the output image.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.color_twist(*inputs, **kwargs)¶

Adjusts hue, saturation, brightness and contrast of the image.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList ('HWC', 'FHWC', 'DHWC')) – Input to the operator.

Keyword Arguments

brightness (float or TensorList of float, optional, default = 1.0) –
Brightness change factor.

Values must be non-negative.

Example values:
- 0 - Black image.
- 1 - No change.
- 2 - Increase brightness twice.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
contrast (float or TensorList of float, optional, default = 1.0) –
Contrast change factor.

Values must be non-negative.

Example values:
- 0 - Uniform grey image.
- 1 - No change.
- 2 - Increase brightness twice.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.UINT8) –
Output data type.

If not set, the input type is used.
hue (float or TensorList of float, optional, default = 0.0) – Hue change, in degrees.
image_type (nvidia.dali.types.DALIImageType, optional, default = DALIImageType.RGB) – The color space of the input and the output image.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
saturation (float or TensorList of float, optional, default = 1.0) –
Saturation change factor.

Values must be non-negative.

Example values:
- 0 - Completely desaturated image.
- 1 - No change to image’s saturation.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.constant(*inputs, **kwargs)¶

Produces a batch of constant tensors.

The floating point input data should be placed in the fdata argument and integer data in idata. The data, which can be a flat vector of values or a scalar, is then reshaped according to the shape argument. If the data is scalar, it will be broadcast to fill the entire shape.

The operator only performs meaningful work at first invocation; subsequent calls will return a reference to the same memory.

The operator can be automatically instantiated in Python with a call to types.Constant(value, dtype, shape, layout). The value can be a scalar, a tuple, a list, or a numpy array. If not explicitly overridden, the shape and dtype, will be taken from the array.

Warning

64-bit integer and double precision arrays are not supported and will be silently downgraded to 32-bit.

Supported backends

‘cpu’
‘gpu’

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.NO_TYPE) –
Output data type.

If this value is not set, the output is float if the fdata argument is used and int if idata is used.
fdata (float or list of float, optional, default = []) –
Contents of the constant that is produced (for floating point types).

Note

fdata and idata are mutually exclusive, and one of them is required.
idata (int or list of int, optional, default = []) –
Contents of the constant that is produced (for integer point types).

Note

fdata and idata are mutually exclusive, and one of them is required.
layout (layout str, optional, default = ‘’) –
Layout info.

If set and not empty, the layout must match the dimensionality of the output.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shape (int or list of int, optional, default = []) – The desired shape of the output. If not set, the data is assumed to be 1D

nvidia.dali.fn.contrast(*inputs, **kwargs)¶

Adjusts the contrast of the images.

The contrast is adjusted based on the following formula:

out = contrast_center + contrast * (in - contrast_center)

This operator can also change the type of data.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList ('FHWC', 'DHWC', 'HWC')) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
contrast (float or TensorList of float, optional, default = 1.0) – The contrast multiplier, where 0.0 produces the uniform grey.
contrast_center (float, optional, default = 0.5) –
The intensity level that is unaffected by contrast.

This is the value that all pixels assume when the contrast is zero. When not set, the half of the input type’s positive range (or 0.5 for float) is used.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.NO_TYPE) –
Output data type.

If not set, the input type is used.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.coord_flip(*inputs, **kwargs)¶

Transforms vectors or points by flipping (reflecting) their coordinates with respect to a given center.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
center_x (float or TensorList of float, optional, default = 0.5) – Flip center in the horizontal axis.
center_y (float or TensorList of float, optional, default = 0.5) – Flip center in the vertical axis.
center_z (float or TensorList of float, optional, default = 0.5) – Flip center in the depthwise axis.
flip_x (int or TensorList of int, optional, default = 1) – Flip the horizontal (x) coordinate.
flip_y (int or TensorList of int, optional, default = 0) – Flip the vertical (y) coordinate.
flip_z (int or TensorList of int, optional, default = 0) – Flip the depthwise (z) coordinate.
layout (layout str, optional, default = ‘’) –
Determines the order of coordinates in the input.

The string should consist of the following characters:
- ”x” (horizontal coordinate),
- ”y” (vertical coordinate),
- ”z” (depthwise coordinate),
Note

If left empty, depending on the number of dimensions, the “x”, “xy”, or “xyz” values are assumed.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.coord_transform(*inputs, **kwargs)¶

Applies a linear transformation to points or vectors.

The transformation has the form:

out = M * in + T

Where M is a m x n matrix and T is a translation vector with m components. Input must consist of n-element vectors or points and the output has m components.

This operator can be used for many operations. Here’s the (incomplete) list:

applying affine transform to point clouds

projecting points onto a subspace

some color space conversions, for example RGB to YCbCr or grayscale

linear operations on colors, like hue rotation, brighness and contrast adjustment

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

M (float or list of float or TensorList of float, optional) –
The matrix used for transforming the input vectors.

If left unspecified, identity matrix is used.

The matrix M does not need to be square - if it’s not, the output vectors will have a number of components equal to the number of rows in M.

If a scalar value is provided, M is assumed to be a square matrix with that value on the diagonal. The size of the matrix is then assumed to match the number of components in the input vectors.
MT (float or list of float or TensorList of float, optional) –
A block matrix [M T] which combines the arguments M and T.

Providing a scalar value for this argument is equivalent to providing the same scalar for M and leaving T unspecified.

The number of columns must be one more than the number of components in the input. This argument is mutually exclusive with M and T.
T (float or list of float or TensorList of float, optional) –
The translation vector.

If left unspecified, no translation is applied unless MT argument is used.

The number of components of this vector must match the number of rows in matrix M. If a scalar value is provided, that value is broadcast to all components of T and the number of components is chosen to match the number of rows in M.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.FLOAT) –
Data type of the output coordinates.

If an integral type is used, the output values are rounded to the nearest integer and clamped to the dynamic range of this type.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.copy(*inputs, **kwargs)¶

Creates a copy of the input tensor.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.crop(*inputs, **kwargs)¶

Crops the images with the specified window dimensions and window position (upper left corner).

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
crop (float or list of float, optional, default = [0.0, 0.0]) –
Shape of the cropped image, specified as a list of values (for example, (crop_H, crop_W) for the 2D crop and (crop_D, crop_H, crop_W) for the volumetric crop).

Providing crop argument is incompatible with providing separate arguments such as crop_d, crop_h, and crop_w.
crop_d (float or TensorList of float, optional, default = 0.0) –
Applies only to volumetric inputs; cropping window depth (in voxels).

crop_w, crop_h, and crop_d must be specified together. Providing values for crop_w, crop_h, and crop_d is incompatible with providing the fixed crop window dimensions (argument crop).
crop_h (float or TensorList of float, optional, default = 0.0) –
Cropping the window height (in pixels).

Providing values for crop_w and crop_h is incompatible with providing fixed crop window dimensions (argument crop).
crop_pos_x (float or TensorList of float, optional, default = 0.5) –
Normalized (0.0 - 1.0) horizontal position of the cropping window (upper left corner).

The actual position is calculated as crop_x = crop_x_norm * (W - crop_W), where crop_x_norm is the normalized position, W is the width of the image, and crop_W is the width of the cropping window.
crop_pos_y (float or TensorList of float, optional, default = 0.5) –
Normalized (0.0 - 1.0) vertical position of the start of the cropping window (typically, the upper left corner).

The actual position is calculated as crop_y = crop_y_norm * (H - crop_H), where crop_y_norm is the normalized position, H is the height of the image, and crop_H is the height of the cropping window.
crop_pos_z (float or TensorList of float, optional, default = 0.5) –
Applies only to volumetric inputs.

Normalized (0.0 - 1.0) normal position of the cropping window (front plane). The actual position is calculated as crop_z = crop_z_norm * (D - crop_D), where crop_z_norm is the normalized position, D is the depth of the image and crop_D is the depth of the cropping window.
crop_w (float or TensorList of float, optional, default = 0.0) –
Cropping window width (in pixels).

Providing values for crop_w and crop_h is incompatible with providing fixed crop window dimensions (argument crop).
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.NO_TYPE) –
Output data type.

Supported types: FLOAT, FLOAT16, and UINT8.

If not set, the input type is used.
fill_values (float or list of float, optional, default = [0.0]) –
Determines padding values and is only relevant if out_of_bounds_policy is set to “pad”.

If a scalar value is provided, it will be used for all the channels. If multiple values are provided, the number of values and channels must be identical (extent of dimension C in the layout) in the output slice.
image_type (nvidia.dali.types.DALIImageType) –

Warning

The argument image_type is no longer used and will be removed in a future release.
out_of_bounds_policy (str, optional, default = ‘error’) –
Determines the policy when slicing the out of bounds area of the input.

Here is a list of the supported values:
- "error" (default): Attempting to slice outside of the bounds of the input will produce an error.
- "pad": The input will be padded as needed with zeros or any other value that is specified with the fill_values argument.
- "trim_to_shape": The slice window will be cut to the bounds of the input.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
output_dtype (nvidia.dali.types.DALIDataType) –

Warning

The argument output_dtype is a deprecated alias for dtype. Use dtype instead.

nvidia.dali.fn.crop_mirror_normalize(*inputs, **kwargs)¶

Performs fused cropping, normalization, format conversion (NHWC to NCHW) if desired, and type casting.

Normalization takes the input images and produces the output by using the following formula:

output = scale * (input - mean) / std + shift

Note

If no cropping arguments are specified, only mirroring and normalization will occur.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
crop (float or list of float, optional, default = [0.0, 0.0]) –
Shape of the cropped image, specified as a list of values (for example, (crop_H, crop_W) for the 2D crop and (crop_D, crop_H, crop_W) for the volumetric crop).

Providing crop argument is incompatible with providing separate arguments such as crop_d, crop_h, and crop_w.
crop_d (float or TensorList of float, optional, default = 0.0) –
Applies only to volumetric inputs; cropping window depth (in voxels).

crop_w, crop_h, and crop_d must be specified together. Providing values for crop_w, crop_h, and crop_d is incompatible with providing the fixed crop window dimensions (argument crop).
crop_h (float or TensorList of float, optional, default = 0.0) –
Cropping the window height (in pixels).

Providing values for crop_w and crop_h is incompatible with providing fixed crop window dimensions (argument crop).
crop_pos_x (float or TensorList of float, optional, default = 0.5) –
Normalized (0.0 - 1.0) horizontal position of the cropping window (upper left corner).

The actual position is calculated as crop_x = crop_x_norm * (W - crop_W), where crop_x_norm is the normalized position, W is the width of the image, and crop_W is the width of the cropping window.
crop_pos_y (float or TensorList of float, optional, default = 0.5) –
Normalized (0.0 - 1.0) vertical position of the start of the cropping window (typically, the upper left corner).

The actual position is calculated as crop_y = crop_y_norm * (H - crop_H), where crop_y_norm is the normalized position, H is the height of the image, and crop_H is the height of the cropping window.
crop_pos_z (float or TensorList of float, optional, default = 0.5) –
Applies only to volumetric inputs.

Normalized (0.0 - 1.0) normal position of the cropping window (front plane). The actual position is calculated as crop_z = crop_z_norm * (D - crop_D), where crop_z_norm is the normalized position, D is the depth of the image and crop_D is the depth of the cropping window.
crop_w (float or TensorList of float, optional, default = 0.0) –
Cropping window width (in pixels).

Providing values for crop_w and crop_h is incompatible with providing fixed crop window dimensions (argument crop).
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.FLOAT) –
Output data type.

Supported types: FLOAT, FLOAT16, INT8, UINT8.
fill_values (float or list of float, optional, default = [0.0]) –
Determines padding values and is only relevant if out_of_bounds_policy is set to “pad”.

If a scalar value is provided, it will be used for all the channels. If multiple values are provided, the number of values and channels must be identical (extent of dimension C in the layout) in the output slice.
image_type (nvidia.dali.types.DALIImageType) –

Warning

The argument image_type is no longer used and will be removed in a future release.
mean (float or list of float or TensorList of float, optional, default = [0.0]) – Mean pixel values for image normalization.
mirror (int or TensorList of int, optional, default = 0) – If nonzero, the image will be flipped (mirrored) horizontally.
out_of_bounds_policy (str, optional, default = ‘error’) –
Determines the policy when slicing the out of bounds area of the input.

Here is a list of the supported values:
- "error" (default): Attempting to slice outside of the bounds of the input will produce an error.
- "pad": The input will be padded as needed with zeros or any other value that is specified with the fill_values argument.
- "trim_to_shape": The slice window will be cut to the bounds of the input.
output_layout (layout str, optional, default = ‘CHW’) – Tensor data layout for the output.
pad_output (bool, optional, default = False) – Determines whether to pad the output to the number of channels as a power of 2).
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
scale (float, optional, default = 1.0) –
The value by which the result is multiplied.

This argument is useful when using integer outputs to improve dynamic range utilization.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shift (float, optional, default = 0.0) –
The value added to the (scaled) result.

This argument is useful when using unsigned integer outputs to improve dynamic range utilization.
std (float or list of float or TensorList of float, optional, default = [1.0]) – Standard deviation values for image normalization.
output_dtype (nvidia.dali.types.DALIDataType) –

Warning

The argument output_dtype is a deprecated alias for dtype. Use dtype instead.

nvidia.dali.fn.dl_tensor_python_function(*inputs, **kwargs)¶

Executes a Python function that operates on DLPack tensors.

The function should not modify input tensors.

For the GPU operator, it is the user’s responsibility to synchronize the device code with DALI. To synchronize the device code with DALI, synchronize DALI’s work before the operator call with the synchronize_stream flag (enabled by default) and ensure that the scheduled device tasks are finished in the operator call. The GPU code can be executed on the CUDA stream used by DALI, which can be obtained by calling the current_dali_stream() function. In this case, the synchronize_stream flag can be set to False.

Warning

This operator is not compatible with TensorFlow integration.

This operator allows sequence inputs and supports volumetric data.

This operator will not be optimized out of the graph.

Supported backends

‘cpu’
‘gpu’

Parameters

input[0..255] (TensorList, optional) – This function accepts up to 256 optional positional inputs

Keyword Arguments

function (object) – Function object.
batch_processing (bool, optional, default = False) –
Determines whether the function is invoked once per batch or separately for every sample in the batch.

If set to True, the function will receive its arguments as lists of DLPack tensors.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
num_outputs (int, optional, default = 1) – Number of outputs.
output_layouts (layout str or list of layout str, optional) –
Tensor data layouts for the outputs.

This argument can be a list that contains a distinct layout for each output. If the list has fewer than num_outputs elements, only the first outputs have the layout set and the rest of the outputs have no layout assigned.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
synchronize_stream (bool, optional, default = True) –
Ensures that DALI synchronizes its CUDA stream before calling the Python function.

Warning

This argument should be set to False only if the called function schedules device work to the stream that is used by DALI.

nvidia.dali.fn.dump_image(*inputs, **kwargs)¶

Save images in batch to disk in PPM format.

Useful for debugging.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
input_layout (layout str, optional, default = ‘HWC’) – Layout of input images.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
suffix (str, optional, default = ‘’) – Suffix to be added to output file names.

nvidia.dali.fn.element_extract(*inputs, **kwargs)¶

Extracts one or more elements from input sequence.

The outputs are slices in the first (outermost) dimension of the input. There are as many outputs as the elements provided in the element_map.

For example, for element_map = [2, 0, 3] there will be three outputs, containing 2nd, 0th and 3rd element of the input sequences respectively.

The input layout, if provided, must begin with F dimension. The outputs will have one less dimension than the input, that is for FHWC inputs, the outputs will be HWC elements.

This operator expects sequence inputs.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

element_map (int or list of int) – Indices of the elements to extract.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.erase(*inputs, **kwargs)¶

Erases one or more regions from the input tensors.

The region is specified by an anchor (starting point) and a shape (dimensions). Only the relevant dimensions are specified. Not specified dimensions are treated as if the entire range of the axis was provided. To specify multiple regions, anchor and shape represent multiple points consecutively (for example, anchor = (y0, x0, y1, x1, …) and shape = (h0, w0, h1, w1, …)). The anchor and shape arguments are interpreted based on the value of the axis_names argument, or, alternatively, the value of the axes argument. If no axis_names or axes arguments are provided, all dimensions except C (channels) must be specified.

Example 1:

anchor = (10, 20), shape = (190, 200), axis_names = “HW”, fill_value = 0

input: layout = “HWC”, shape = (300, 300, 3)

The erase region covers the range between 10 and 200 in the vertical dimension (height) and between 20 and 220 in the horizontal dimension (width). The range for the channel dimension was not specified, so it is between 0 and 3. What gives:

output[y, x, c] = 0   if 20 <= x < 220 and 10 <= y < 200
output[y, x, c] = input[y, x, c]  otherwise

Example 2:

anchor = (10, 250), shape = (20, 30), axis_names = “W”, fill_value = (118, 185, 0)

input: layout = “HWC”, shape = (300, 300, 3)

Two erase regions are provided, which covers two vertical bands that range from x=(10, 30) and x=(250, 280), respectively. Each pixel in the erased regions is filled with a multi-channel value (118, 185, 0). What gives:

output[y, x, :] = (118, 185, 0)   if 10 <= x < 30 or 250 <= x < 280
output[y, x, :] = input[y, x, :]  otherwise

Example 3:

anchor = (0.15, 0.15), shape = (0.3, 0.3), axis_names = “HW”, fill_value = 100, normalized = True

input: layout = “HWC”, shape = (300, 300, 3)

One erase region with normalized coordinates in the height, and the width dimensions is provided. A fill value is provided for all the channels. The coordinates can be transformed to the absolute by multiplying by the input shape. What gives:

output[y, x, c] = 100             if 0.15 * 300 <= x < (0.3 + 0.15) * 300 and 0.15 * 300 <= y < (0.3 + 0.15) * 300
output[y, x, c] = input[y, x, c]  otherwise

Example 4: anchor = (0.15, 0.15), shape = (20, 30), normalized_anchor = True, normalized_shape = False

input: layout = “HWC”, shape = (300, 300, 3)

One erase region with an anchor is specified in normalized coordinates and the shape in absolute coordinates. Since no axis_names is provided, the anchor and shape must contain all dimensions except “C” (channels). What gives:

output[y, x, c] = 0               if 0.15 * 300 <= x < (0.15 * 300) + 20 and (0.15 * 300) <= y < (0.15 * 300) + 30
output[y, x, c] = input[y, x, c]  otherwise

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

anchor (float or list of float or TensorList of float, optional, default = []) –
Coordinates for the anchor or the starting point of the erase region.

Only the coordinates of the relevant dimensions that are specified by axis_names or axes should be provided.
axes (int or list of int, optional, default = [1, 0]) –
Order of dimensions used for anchor and shape arguments, as dimension indices.

For instance, axes=(1, 0) means the coordinates in anchor and shape refer to axes 1 and 0 in that particular order.
axis_names (str, optional, default = ‘HW’) –
Order of dimensions that are used for the anchor and shape arguments, as described in the layout.

For instance, axis_names=”HW” means that the coordinates in anchor and shape refer to dimensions H (height) and W (width) in that particular order.

Note

axis_name*s has a higher priority than *axes.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
centered_anchor (bool, optional, default = False) –
If set to True, the anchors refer to the center of the region instead of the top-left corner.

This results in centered erased regions at the specified anchor.
fill_value (float or list of float, optional, default = [0.0]) –
Value to fill the erased region.

Might be specified as one value (for example, 0) or a multi-channel value (for example, (200, 210, 220)). If a multi-channel fill value is provided, the input layout should contain a channel dimension C.
normalized (bool, optional, default = False) –
Determines whether the anchor and shape arguments should be interpreted as normalized (range [0.0, 1.0]) or absolute coordinates.

Providing a value for the normalized_shape and normalized_anchor arguments separately is mutually exclusive.
normalized_anchor (bool, optional, default = False) –
Determines whether the anchor argument should be interpreted as normalized (range [0.0, 1.0]) or absolute coordinates.

Providing a value for normalized is mutually exclusive.
normalized_shape (bool, optional, default = False) –
Determines whether the shape argument should be interpreted as normalized (range [0.0, 1.0]) or absolute coordinates.

Providing a value for normalized is mutually exclusive.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shape (float or list of float or TensorList of float, optional, default = []) –
Values for shape or dimensions of the erase region.

Only the coordinates of the relevant dimensions that are specified by axis_names or axes should be provided.

nvidia.dali.fn.expand_dims(*inputs, **kwargs)¶

Insert new dimension(s) with extent 1 to the data shape.

The new dimensions are inserted at the positions specified by axes.

If new_axis_names is provided, the new dimension names will be inserted in the data layout, at the positions specified by axes. If new_axis_names is not provided, the output data layout will be empty.”

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

data (TensorList) – Data to be expanded

Keyword Arguments

axes (int or list of int or TensorList of int) – Indices at which the new dimensions are inserted.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
new_axis_names (layout str, optional, default = ‘’) –
Names of the new dimensions in the data layout.

The length of new_axis_names must match the length of axes. If argument isn’t be provided, the layout will be cleared.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.external_source(source=None, num_outputs=None, *, cycle=None, name=None, device='cpu', layout=None, dtype=None, cuda_stream=None, use_copy_kernel=None, batch=True, **kwargs)¶

Creates a data node which is populated with data from a Python source. The data can be provided by the source function or iterable, or it can be provided by pipeline.feed_input(name, data, layout, cuda_stream) inside pipeline.iter_setup.

In the case of the GPU input, it is the user responsibility to modify the provided GPU memory content only using provided stream (DALI schedules a copy on it and all work is properly queued). If no stream is provided feeding input blocks until the provided memory is copied to the internal buffer.

Note

nvidia.dali.fn.external_source() operator is partially compatible with TensorFlow integration via nvidia.dali.plugin.tf.experimental.DALIDatasetWithInputs(). Please refer to its documentation for details.

Note

To return a batch of copies of the same tensor, use nvidia.dali.types.Constant(), which is more performant.

Parameters

source (callable or iterable) –
The source of the data.

The source is polled for data (via a call source() or next(source)) when the pipeline needs input for the next iteration. Depending on the value of num_outputs, the source can supply one or more data items. The data item can be a whole batch (default) or a single batch entry (when batch==False). If num_outputs is not set, the source is expected to return one item (a batch or a sample). If this value is specified (even if its value is 1), the data is expected to a be tuple, or list, where each element corresponds to respective return value of the external_source.

The data samples must be in one of the compatible array types:
- NumPy ndarray (CPU)
- MXNet ndarray (CPU)
- PyTorch tensor (CPU or GPU)
- CuPy array (GPU)
- objects implementing __cuda_array_interface__
- DALI Tensor object
Batch sources must produce entire batches of data. This can be achieved either by adding a new outermost dimension to an array or by returning a list of arrays (in which case they can be of different size, but must have the same rank and element type). A batch source can also produce a DALI TensorList object, which can be an output of another DALI pipeline.

A per-batch source may accept one positional argument. If it does, it is the index of current iteration within epoch and consecutive calls will be source(0), source(1), and so on. If batch_info is set to True, instance of nvidia.dali.types.BatchInfo will be passed to the source, instead of a plain index.

A per-sample source may accept one positional argument of type nvidia.dali.types.SampleInfo, which contains index of the sample in current epoch and in the batch, as well as current iteration number.

If the source is a generator function, the function is invoked and treated as an iterable. However, unlike a generator, the function can be used with cycle. In this case, the function will be called again when the generator reaches the end of iteration.

For GPU inputs, it is a user’s responsibility to modify the provided GPU memory content only in the provided stream. DALI schedules a copy on this stream, and all work is properly queued. If no stream is provided, DALI will use a default, with a best-effort approach at correctness. See the cuda_stream argument documentation for more information.
num_outputs (int, optional) –
If specified, denotes the number of TensorLists that are produced by the source function.

If set, the operator returns a list of DataNode objects, otherwise a single DataNode object is returned.

Keyword Arguments

cycle (string or bool, optional) –
Specifies if and how to cycle through the source. It can be one of the following values:
- "no", False or None - don’t cycle; StopIteration is raised whe end of data is reached; this is the default behavior
- "quiet" or True - the data is repeated indefinitely,
- "raise" - when the end of data is reached, StopIteration is raised, but the iteration is restarted on subsequent call.
This flag requires that the source is a collection, for example, an iterable object where iter(source) returns a fresh iterator on each call, or a generator function. In the latter case, the generator function is called again when more data than was yielded by the function is requested.

Specifying "raise" can be used with DALI iterators to create a notion of epoch.
name (str, optional) –
The name of the data node.

Used when feeding the data in iter_setup and can be omitted if the data is provided by source.
layout (layout str or list/tuple thereof, optional) –
If provided, sets the layout of the data.

When num_outputs > 1, the layout can be a list that contains a distinct layout for each output. If the list has fewer than num_outputs elements, only the first outputs have the layout set, the rest of the outputs don’t have a layout set.
dtype (nvidia.dali.types.DALIDataType or list/tuple thereof, optional) –
Input data type.

The operator will validate that the fetched data is of the provided type. If the argument is omitted or DALIDataType.NO_TYPE is passed, the operator will infer the type from the provided data.

This argument will be required starting from DALI 2.0.
cuda_stream (optional, cudaStream_t or an object convertible to cudaStream_t, such as cupy.cuda.Stream or torch.cuda.Stream) –
The CUDA stream is used to copy data to the GPU or from a GPU source.

If this parameter is not set, a best-effort will be taken to maintain correctness. That is, if the data is provided as a tensor/array from a recognized library such as CuPy or PyTorch, the library’s current stream is used. Although this approach works in typical scenarios, with advanced use cases, and code that uses unsupported libraries, you might need to explicitly supply the stream handle.

This argument has two special values:
- 0 - Use the default CUDA stream
- 1 - Use DALI’s internal stream
If internal stream is used, the call to feed_input will block until the copy to internal buffer is complete, since there’s no way to synchronize with this stream to prevent overwriting the array with new data in another stream.
use_copy_kernel (bool, optional) –
If set to True, DALI will use a CUDA kernel to feed the data instead of cudaMemcpyAsync (default).

Note

This is applicable only when copying data to and from GPU memory.
blocking (bool, optional) – Determines whether the external source should wait until data is available or just fail when the data is not available.
no_copy (bool, optional) –
Determines whether DALI should copy the buffer when feed_input is called.

If set to True, DALI passes the user memory directly to the pipeline, instead of copying it. It is the user responsibility to keep the buffer alive and unmodified until it is consumed by the pipeline.

The buffer can be modified or freed again after the output of the relevant iterations has been consumed. Effectively, it happens after Pipeline’s prefetch_queue_depth or cpu_queue_depth * gpu_queue_depth (when they are not equal) iterations following the feed_input call.

The memory location must match the specified device parameter of the operator. For the CPU, the provided memory can be one contiguous buffer or a list of contiguous Tensors. For the GPU, to avoid extra copy, the provided buffer must be contiguous. If you provide a list of separate Tensors, there will be an additional copy made internally, consuming both memory and bandwidth.

Automatically set to True when parallel=True
batch (bool, optional) –
If set to True or None, the source is expected to produce an entire batch at once. If set to False, the source is called per-sample.

Setting parallel to True automatically sets batch to False if it was not provided.
batch_info (bool, optional, default = False) – Controls if a callable source that accepts an argument and returns batches should receive BatchInfo instance or just an integer representing the iteration number. If set to False (the default), only the integer is passed. If source is not callable, does not accept arguments or batch is set to False, setting this flag has no effect.
parallel (bool, optional, default = False) –
If set to True, the corresponding pipeline will start a pool of Python workers to run the callback in parallel. You can specify the number of workers by passing py_num_workers into pipeline’s constructor.

When parallel is set to True, samples returned by source must be NumPy/MXNet/PyTorch CPU arrays or TensorCPU instances.

Acceptable sources depend on the value specified for batch parameter.

If batch is set to False, the source must be:
- a callable (a function or an object with __call__ method) that accepts exactly one argument (SampleInfo instance that represents the index of the requested sample).
If batch is set to True, the source can be either:
- a callable that accepts exactly one argument (either BatchInfo instance or an integer - see batch_info for details)
- an iterable,
- a generator function.
Warning

Irrespective of batch value, callables should be stateless - they should produce requested sample or batch solely based on the SampleInfo/BatchInfo instance or index in batch, so that they can be run in parallel in a number of workers.

The source callback must raise a StopIteration when the end of the data is reached. Note, that due to prefetching, the callback may be invoked with a few iterations past the end of dataset - make sure it consistently raises a StopIteration in that case.

Note

Callable source can be run in parallel by multiple workers. For batch=True multiple batches can be prepared in parallel, with batch=False it is possible to parallelize computation within the batch.

When batch=True, callables performance might especially benefit from increasing prefetch_queue_depth so that a few next batches can be computed in parallel.

Note

Iterator or generator function will be assigned to a single worker that will iterate over them. The main advantage is execution in parallel to the main Python process, but due to their state it is not possible to calculate more than one batch at a time.
prefetch_queue_depth (int, option, default = 1) – When run in parallel=True mode, specifies the number of batches to be computed in advance and stored in the internal buffer, otherwise parameter is ignored.

nvidia.dali.fn.fast_resize_crop_mirror(*inputs, **kwargs)¶

Performs a fused resize, crop, mirror operation.

The operator handles both fixed and random resizing and cropping. Backprojects the desired crop through the resize operation to reduce the amount of work that is performed.

Supported backends

‘cpu’

Parameters

input (TensorList ('HWC')) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
crop (float or list of float, optional, default = [0.0, 0.0]) –
Shape of the cropped image, specified as a list of values (for example, (crop_H, crop_W) for the 2D crop and (crop_D, crop_H, crop_W) for the volumetric crop).

Providing crop argument is incompatible with providing separate arguments such as crop_d, crop_h, and crop_w.
crop_d (float or TensorList of float, optional, default = 0.0) –
Applies only to volumetric inputs; cropping window depth (in voxels).

crop_w, crop_h, and crop_d must be specified together. Providing values for crop_w, crop_h, and crop_d is incompatible with providing the fixed crop window dimensions (argument crop).
crop_h (float or TensorList of float, optional, default = 0.0) –
Cropping the window height (in pixels).

Providing values for crop_w and crop_h is incompatible with providing fixed crop window dimensions (argument crop).
crop_pos_x (float or TensorList of float, optional, default = 0.5) –
Normalized (0.0 - 1.0) horizontal position of the cropping window (upper left corner).

The actual position is calculated as crop_x = crop_x_norm * (W - crop_W), where crop_x_norm is the normalized position, W is the width of the image, and crop_W is the width of the cropping window.
crop_pos_y (float or TensorList of float, optional, default = 0.5) –
Normalized (0.0 - 1.0) vertical position of the start of the cropping window (typically, the upper left corner).

The actual position is calculated as crop_y = crop_y_norm * (H - crop_H), where crop_y_norm is the normalized position, H is the height of the image, and crop_H is the height of the cropping window.
crop_pos_z (float or TensorList of float, optional, default = 0.5) –
Applies only to volumetric inputs.

Normalized (0.0 - 1.0) normal position of the cropping window (front plane). The actual position is calculated as crop_z = crop_z_norm * (D - crop_D), where crop_z_norm is the normalized position, D is the depth of the image and crop_D is the depth of the cropping window.
crop_w (float or TensorList of float, optional, default = 0.0) –
Cropping window width (in pixels).

Providing values for crop_w and crop_h is incompatible with providing fixed crop window dimensions (argument crop).
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.NO_TYPE) –
Output data type.

Supported types: FLOAT, FLOAT16, and UINT8.

If not set, the input type is used.
fill_values (float or list of float, optional, default = [0.0]) –
Determines padding values and is only relevant if out_of_bounds_policy is set to “pad”.

If a scalar value is provided, it will be used for all the channels. If multiple values are provided, the number of values and channels must be identical (extent of dimension C in the layout) in the output slice.
image_type (nvidia.dali.types.DALIImageType) –

Warning

The argument image_type is no longer used and will be removed in a future release.
interp_type (nvidia.dali.types.DALIInterpType, optional, default = DALIInterpType.INTERP_LINEAR) – Type of interpolation used.
max_size (float or list of float, optional) –
Limit of the output size.

When the operator is configured to keep aspect ratio and only the smaller dimension is specified, the other(s) can grow very large. This can happen when using resize_shorter argument or “not_smaller” mode or when some extents are left unspecified.

This parameter puts a limit to how big the output can become. This value can be specified per-axis or uniformly for all axes.

Note

When used with “not_smaller” mode or resize_shorter argument, max_size takes precedence and the aspect ratio is kept - for example, resizing with mode="not_smaller", size=800, max_size=1400 an image of size 1200x600 would be resized to 1400x700.
mirror (int or TensorList of int, optional, default = 0) –
Mask for the horizontal flip.

Supported values:
- 0 - Do not perform horizontal flip for this image.
- 1 - Performs horizontal flip for this image.
mode (str, optional, default = ‘default’) –
Resize mode.

Here is a list of supported modes:
- "default" - image is resized to the specified size.
  
  Missing extents are scaled with the average scale of the provided ones.
- "stretch" - image is resized to the specified size.
  
  Missing extents are not scaled at all.
- "not_larger" - image is resized, keeping the aspect ratio, so that no extent of the output image exceeds the specified size.
  
  For example, a 1280x720, with a desired output size of 640x480, actually produces a 640x360 output.
- "not_smaller" - image is resized, keeping the aspect ratio, so that no extent of the output image is smaller than specified.
  
  For example, a 640x480 image with a desired output size of 1920x1080, actually produces a 1920x1440 output.
  
  This argument is mutually exclusive with resize_longer and resize_shorter
out_of_bounds_policy (str, optional, default = ‘error’) –
Determines the policy when slicing the out of bounds area of the input.

Here is a list of the supported values:
- "error" (default): Attempting to slice outside of the bounds of the input will produce an error.
- "pad": The input will be padded as needed with zeros or any other value that is specified with the fill_values argument.
- "trim_to_shape": The slice window will be cut to the bounds of the input.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
resize_longer (float or TensorList of float, optional, default = 0.0) –
The length of the longer dimension of the resized image.

This option is mutually exclusive with resize_shorter and explicit size arguments, and the operator keeps the aspect ratio of the original image. This option is equivalent to specifying the same size for all dimensions and mode="not_larger".
resize_shorter (float or TensorList of float, optional, default = 0.0) –
The length of the shorter dimension of the resized image.

This option is mutually exclusive with resize_longer and explicit size arguments, and the operator keeps the aspect ratio of the original image. This option is equivalent to specifying the same size for all dimensions and mode="not_smaller". The longer dimension can be bounded by setting the max_size argument. See max_size argument doc for more info.
resize_x (float or TensorList of float, optional, default = 0.0) –
The length of the X dimension of the resized image.

This option is mutually exclusive with resize_shorter, resize_longer and size. If the resize_y is unspecified or 0, the operator keeps the aspect ratio of the original image. A negative value flips the image.
resize_y (float or TensorList of float, optional, default = 0.0) –
The length of the Y dimension of the resized image.

This option is mutually exclusive with resize_shorter, resize_longer and size. If the resize_x is unspecified or 0, the operator keeps the aspect ratio of the original image. A negative value flips the image.
resize_z (float or TensorList of float, optional, default = 0.0) –
The length of the Z dimension of the resized volume.

This option is mutually exclusive with resize_shorter, resize_longer and size. If the resize_x and resize_y are left unspecified or 0, then the op will keep the aspect ratio of the original volume. Negative value flips the volume.
roi_end (float or list of float or TensorList of float, optional) –
End of the input region of interest (ROI).

Must be specified together with roi_start. The coordinates follow the tensor shape order, which is the same as size. The coordinates can be either absolute (in pixels, which is the default) or relative (0..1), depending on the value of relative_roi argument. If the ROI origin is greater than the ROI end in any dimension, the region is flipped in that dimension.
roi_relative (bool, optional, default = False) – If true, ROI coordinates are relative to the input size, where 0 denotes top/left and 1 denotes bottom/right
roi_start (float or list of float or TensorList of float, optional) –
Origin of the input region of interest (ROI).

Must be specified together with roi_end. The coordinates follow the tensor shape order, which is the same as size. The coordinates can be either absolute (in pixels, which is the default) or relative (0..1), depending on the value of relative_roi argument. If the ROI origin is greater than the ROI end in any dimension, the region is flipped in that dimension.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
size (float or list of float or TensorList of float, optional) –
The desired output size.

Must be a list/tuple with one entry per spatial dimension, excluding video frames and channels. Dimensions with a 0 extent are treated as absent, and the output size will be calculated based on other extents and mode argument.
subpixel_scale (bool, optional, default = True) –
If True, fractional sizes, directly specified or calculated, will cause the input ROI to be adjusted to keep the scale factor.

Otherwise, the scale factor will be adjusted so that the source image maps to the rounded output size.
output_dtype (nvidia.dali.types.DALIDataType) –

Warning

The argument output_dtype is a deprecated alias for dtype. Use dtype instead.

nvidia.dali.fn.file_reader(*inputs, **kwargs)¶: Warning

This operator is now deprecated. Use readers.file() instead.

In DALI 1.0 all readers were moved into a dedicated readers submodule and renamed to follow a common pattern. This is a placeholder operator with identical functionality to allow for backward compatibility.

Legacy alias for readers.file().

nvidia.dali.fn.flip(*inputs, **kwargs)¶

Flips the images in selected dimensions (horizontal, vertical, and depthwise).

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList ('FDHWC', 'FHWC', 'DHWC', 'HWC', 'FCDHW', 'FCHW', 'CDHW', 'CHW')) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
depthwise (int or TensorList of int, optional, default = 0) – Flip the depthwise dimension.
horizontal (int or TensorList of int, optional, default = 1) – Flip the horizontal dimension.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
vertical (int or TensorList of int, optional, default = 0) – Flip the vertical dimension.

nvidia.dali.fn.gaussian_blur(*inputs, **kwargs)¶

Applies a Gaussian Blur to the input.

Gaussian blur is calculated by applying a convolution with a Gaussian kernel, which can be parameterized with windows_size and sigma. If only the sigma is specified, the radius of the Gaussian kernel defaults to ceil(3 * sigma), so the kernel window size is 2 * ceil(3 * sigma) + 1.

If only the window size is provided, the sigma is calculated by using the following formula:

radius = (window_size - 1) / 2
sigma = (radius - 1) * 0.3 + 0.8

The sigma and kernel window size can be specified as one value for all data axes or a value per data axis.

When specifying the sigma or window size per axis, the axes are provided same as layouts, from outermost to innermost.

Note

The channel C and frame F dimensions are not considered data axes. If channels are present, only channel-first or channel-last inputs are supported.

For example, with HWC input, you can provide sigma=1.0 or sigma=(1.0, 2.0) because there are two data axes, H and W.

The same input can be provided as per-sample tensors.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.NO_TYPE) –
Output data type.

Supported type: FLOAT. If not set, the input type is used.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
sigma (float or list of float or TensorList of float, optional, default = [0.0]) – Sigma value for the Gaussian Kernel.
window_size (int or list of int or TensorList of int, optional, default = [0]) – The diameter of the kernel.

nvidia.dali.fn.get_property(*inputs, **kwargs)¶

Returns a property of the tensor passed as an input.

The type of the output will depend on the key of the requested property.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

key (str) –
Specifies, which property is requested.

The following properties are supported:
- "source_info": Returned type: byte-array.
  String-like byte array, which contains information about the origin of the sample. Fox example, fn.get_property() called on tensor loaded via fn.readers.file() returns full path of the file, from which the tensor comes from.
- "layout": Returned type: byte-array
  Data layout in the given Tensor.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.grid_mask(*inputs, **kwargs)¶

Performs the gridmask augumentation (https://arxiv.org/abs/2001.04086).

Zeroes out pixels of an image in a grid-like fashion. The grid consists of squares repeating in x and y directions, with the same spacing in both directions. Can be rotated around the origin.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

angle (float or TensorList of float, optional, default = 0.0) – Angle, in radians, by which the grid is rotated.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
ratio (float or TensorList of float, optional, default = 0.5) – The ratio between black square width and tile width.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shift_x (float or TensorList of float, optional, default = 0.0) – The x component of the translation vector, applied after rotation.
shift_y (float or TensorList of float, optional, default = 0.0) – The y component of the translation vector, applied after rotation.
tile (int or TensorList of int, optional, default = 100) – The length of a single tile, which is equal to width of black squares plus the spacing between them.

nvidia.dali.fn.hsv(*inputs, **kwargs)¶

Adjusts hue, saturation and value (brightness) of the images.

To change the hue, the saturation, and/or the value of the image, pass the corresponding coefficients. Remember that the hue is an additive delta argument, while for saturation and value, the arguments are multiplicative.

This operator accepts images in the RGB color space.

For performance reasons, the operation is approximated by a linear transform in the RGB space. The color vector is projected along the neutral (gray) axis, rotated based on the hue delta, scaled based on the value and saturation multipliers, and restored to the original color space.

This operator allows sequence inputs.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList ('HWC', 'FHWC', 'DHWC')) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.UINT8) –
The output data type.

If a value is not set, the input type is used.
hue (float or TensorList of float, optional, default = 0.0) –
Hue delta, in degrees.

The hue component can be interpreted as an angle and values outside 0-360 range wrap around, as they would in case of rotation.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
saturation (float or TensorList of float, optional, default = 1.0) – The saturation multiplier.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
value (float or TensorList of float, optional, default = 1.0) – The value multiplier.

nvidia.dali.fn.hue(*inputs, **kwargs)¶

Changes the hue level of the image.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList ('HWC', 'FHWC', 'DHWC')) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.UINT8) –
Output data type.

If not set, the input type is used.
hue (float or TensorList of float, optional, default = 0.0) – The hue change in degrees.
image_type (nvidia.dali.types.DALIImageType, optional, default = DALIImageType.RGB) – The color space of the input and the output image.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.image_decoder(*inputs, **kwargs)¶: Warning

This operator is now deprecated. Use decoders.image() instead.

In DALI 1.0 all decoders were moved into a dedicated decoders submodule and renamed to follow a common pattern. This is a placeholder operator with identical functionality to allow for backward compatibility.

Legacy alias for decoders.image().

nvidia.dali.fn.image_decoder_crop(*inputs, **kwargs)¶: Warning

This operator is now deprecated. Use decoders.image_crop() instead.

In DALI 1.0 all decoders were moved into a dedicated decoders submodule and renamed to follow a common pattern. This is a placeholder operator with identical functionality to allow for backward compatibility.

Legacy alias for decoders.image_crop().

nvidia.dali.fn.image_decoder_random_crop(*inputs, **kwargs)¶: Warning

This operator is now deprecated. Use decoders.image_random_crop() instead.

In DALI 1.0 all decoders were moved into a dedicated decoders submodule and renamed to follow a common pattern. This is a placeholder operator with identical functionality to allow for backward compatibility.

Legacy alias for decoders.image_random_crop().

nvidia.dali.fn.image_decoder_slice(*inputs, **kwargs)¶: Warning

This operator is now deprecated. Use decoders.image_slice() instead.

In DALI 1.0 all decoders were moved into a dedicated decoders submodule and renamed to follow a common pattern. This is a placeholder operator with identical functionality to allow for backward compatibility.

Legacy alias for decoders.image_slice().

nvidia.dali.fn.jitter(*inputs, **kwargs)¶

Performs a random Jitter augmentation.

The output images are produced by moving each pixel by a random amount, in the x and y dimensions, and bounded by half of the nDegree parameter.

Supported backends

‘gpu’

Parameters

input (TensorList ('HWC')) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
fill_value (float, optional, default = 0.0) – Color value that is used for padding.
interp_type (nvidia.dali.types.DALIInterpType, optional, default = DALIInterpType.INTERP_NN) – Type of interpolation used.
mask (int or TensorList of int, optional, default = 1) –
Determines whether to apply this augmentation to the input image.

Here are the values:
- 0: Do not apply this transformation.
- 1: Apply this transformation.
nDegree (int, optional, default = 2) – Each pixel is moved by a random amount in the [-nDegree/2, nDegree/2] range
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.jpeg_compression_distortion(*inputs, **kwargs)¶

Introduces JPEG compression artifacts to RGB images.

JPEG is a lossy compression format which exploits characteristics of natural images and human visual system to achieve high compression ratios. The information loss originates from sampling the color information at a lower spatial resolution than the brightness and from representing high frequency components of the image with a lower effective bit depth. The conversion to frequency domain and quantization is applied independently to 8x8 pixel blocks, which introduces additional artifacts at block boundaries.

This operation produces images by subjecting the input to a transformation that mimics JPEG compression with given quality factor followed by decompression.

This operator allows sequence inputs.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList ('HWC', 'FHWC')) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
quality (int or TensorList of int, optional, default = 50) –
JPEG compression quality from 1 (lowest quality) to 100 (highest quality).

Any values outside the range 1-100 will be clamped.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.laplacian(*inputs, **kwargs)¶

Computes the Laplacian of an input.

The Laplacian is calculated as the sum of second order partial derivatives with respect to each spatial dimension. Each partial derivative is approximated with a separable convolution, that uses a derivative window in the direction of the partial derivative and smoothing windows in the remaining axes.

By default, each partial derivative is approximated by convolving along all spatial axes: the axis in partial derivative direction uses derivative window of window_size and the remaining axes are convolved with smoothing windows of the same size. If smoothing_size is specified, the smoothing windows applied to a given axis can have different size than the derivative window. Specifying smoothing_size = 1 implies no smoothing in axes perpendicular to the derivative direction.

Both window_size and smoothing_size can be specified as a single value or per axis. For example, for volumetric input, if window_size=[dz, dy, dx] and smoothing_size=[sz, sy, sx] are specified, the following windows will be used:

for partial derivative in z direction: derivative windows of size dz along z axis, and smoothing windows of size sy and sx along y and x respectively.

for partial derivative in y direction: derivative windows of size dy along y axis, and smoothing windows of size sz and sx along z and x respectively.

for partial derivative in x direction: derivative windows of size dx along x axis, and smoothing windows of size sz and sy along z and y respectively.

Window sizes and smoothing sizes must be odd. The size of a derivative window must be at least 3. Smoothing window can be of size 1, which implies no smoothing along corresponding axis.

To normalize partial derivatives, normalized_kernel=True can be used. Each partial derivative is scaled by 2^(-s + n + 2), where s is the sum of the window sizes used to calculate a given partial derivative (including the smoothing windows) and n is the number of data dimensions/axes. Alternatively, you can specify scale argument to customize scaling factors. Scale can be either a single value or n values, one for every partial derivative.

Operator uses 32-bit floats as an intermediate type.

Note

The channel C and frame F dimensions are not considered data axes. If channels are present, only channel-first or channel-last inputs are supported.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.NO_TYPE) –
Output data type.

Supported type: FLOAT. If not set, the input type is used.
normalized_kernel (bool, optional, default = False) – If set to True, automatically scales partial derivatives kernels. Must be False if scale is specified.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
scale (float or list of float or TensorList of float, optional, default = [1.0]) – Factors to manually scale partial derivatives.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
smoothing_size (int or list of int or TensorList of int, optional) –
Size of smoothing window used in convolutions.

Smoothing size must be odd and between 1 and 23.
window_size (int or list of int or TensorList of int, optional, default = [3]) –
Size of derivative window used in convolutions.

Window size must be odd and between 3 and 23.

nvidia.dali.fn.lookup_table(*inputs, **kwargs)¶

Maps the input to output by using a lookup table that is specified by keys and values, and a default_value for unspecified keys.

For example when keys and values are used to define the lookup table in the following way:

keys[] =   {0,     2,   3,   4,   5,    3}
values[] = {0.2, 0.4, 0.5, 0.6, 0.7, 0.10}
default_value = 0.99

0 <= i < max(keys)
lut[i] = values[keys.index[i]]   if i in keys
lut[i] = default_value           otherwise

the operator creates the following table:

lut[] = {0.2, 0.99, 0.4, 0.10, 0.6, 0.7}  // only last occurrence of a key is considered

and produces the output according to this formula:

Output[i] = lut[Input[i]]   if 0 <= Input[i] <= len(lut)
Output[i] = default_value   otherwise

Here is a practical example, considering the table defined above:

Input[] =  {1,      4,    1,   0,  100,   2,     3,   4}
Output[] = {0.99, 0.6, 0.99, 0.2, 0.99, 0.4,  0.10, 0.6}

Note

Only integer types can be used as inputs for this operator.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
default_value (float, optional, default = 0.0) – Default output value for keys that are not present in the table.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.FLOAT) – Output data type.
keys (int or list of int, optional, default = []) –
A list of input values (keys) in the lookup table.

The length of keys and values argument must match. The values in keys should be in the [0, 65535 ] range.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
values (float or list of float, optional, default = []) –
A list of mapped output values for each keys entry.

The length of the keys and the values argument must match.
output_dtype (nvidia.dali.types.DALIDataType) –

Warning

The argument output_dtype is a deprecated alias for dtype. Use dtype instead.

nvidia.dali.fn.mel_filter_bank(*inputs, **kwargs)¶

Converts a spectrogram to a mel spectrogram by applying a bank of triangular filters.

The frequency (‘f’) dimension is selected from the input layout. In case of no layout, “f”, “ft”, or “*ft” is assumed, depending on the number of dimensions.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
freq_high (float, optional, default = 0.0) –
The maximum frequency.

If this value is not provided, sample_rate/2 is used.
freq_low (float, optional, default = 0.0) – The minimum frequency.
mel_formula (str, optional, default = ‘slaney’) –
Determines the formula that will be used to convert frequencies from hertz to mel and from mel to hertz.

The mel scale is a perceptual scale of pitches, so there is no single formula.

The supported values are:
- slaney, which follows Slaney’s MATLAB Auditory Modelling Work behavior.
  
  This formula is linear under 1 KHz and logarithmic above this value. The implementation is consistent with Librosa’s default implementation.
- htk, which follows O’Shaughnessy’s book formula, m = 2595 * log10(1 + (f/700)).
  
  This value is consistent with the implementation of the Hidden Markov Toolkit (HTK).
nfilter (int, optional, default = 128) – Number of mel filters.
normalize (bool, optional, default = True) –
Determines whether to normalize the triangular filter weights by the width of their frequency bands.
- If set to True, the integral of the filter function is 1.
- If set to False, the peak of the filter function will be 1.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
sample_rate (float, optional, default = 44100.0) – Sampling rate of the audio signal.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.mfcc(*inputs, **kwargs)¶

Computes Mel Frequency Cepstral Coefficiencs (MFCC) from a mel spectrogram.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

axis (int, optional, default = 0) –
Axis over which the transform will be applied.

If a value is not provided, the outer-most dimension will be used.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dct_type (int, optional, default = 2) –
Discrete Cosine Transform type.

The supported types are 1, 2, 3, 4. The formulas that are used to calculate the DCT are equivalent to those described in https://en.wikipedia.org/wiki/Discrete_cosine_transform (the numbers correspond to types listed in https://en.wikipedia.org/wiki/Discrete_cosine_transform#Formal_definition).
lifter (float, optional, default = 0.0) –
Cepstral filtering coefficient, which is also known as the liftering coefficient.

If the lifter coefficient is greater than 0, the MFCCs will be scaled based on the following formula:
```
MFFC[i] = MFCC[i] * (1 + sin(pi * (i + 1) / lifter)) * (lifter / 2)
```
n_mfcc (int, optional, default = 20) – Number of MFCC coefficients.
normalize (bool, optional, default = False) –
If set to True, the DCT uses an ortho-normal basis.

Note

Normalization is not supported when dct_type=1.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.multi_paste(*inputs, **kwargs)¶

Performs multiple pastes from image batch to each of outputs

This operator can also change the type of data.

Supported backends

‘cpu’
‘gpu’

Parameters

images (3D TensorList) –

Batch of input images.

Assumes HWC layout.

Keyword Arguments

in_ids (int or list of int or TensorList of int) – Indices of the inputs to paste data from.
output_size (int or list of int or TensorList of int) – Shape of the output.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.NO_TYPE) – Output data type. If not set, the input type is used.
in_anchors (int or TensorList of int, optional) –
Absolute coordinates of LU corner of the source region.

The anchors are represented as 2D tensors where the first dimension corresponds to the elements of in_ids and the second one is equal to the number of dimensions of the data, excluding channels.

If not provided, all anchors are zero.
out_anchors (int or TensorList of int, optional) –
Absolute coordinates of LU corner of the destination region.

The anchors are represented as 2D tensors where the first dimension corresponds to the elements of in_ids and the second one is equal to the number of dimensions of the data, excluding channels.

If not provided, all anchors are zero.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shapes (int or TensorList of int, optional) –
Shape of the paste regions.

The shapes are represented as 2D tensors where the first dimension corresponds to the elements of in_ids and the second one is equal to the number of dimensions of the data, excluding channels.

If not provided, the shape is calculated so that the region goes from the region anchor
in the input image until the end of the input image.

nvidia.dali.fn.mxnet_reader(*inputs, **kwargs)¶: Warning

This operator is now deprecated. Use readers.mxnet() instead.

In DALI 1.0 all readers were moved into a dedicated readers submodule and renamed to follow a common pattern. This is a placeholder operator with identical functionality to allow for backward compatibility.

Legacy alias for readers.mxnet().

nvidia.dali.fn.nemo_asr_reader(*inputs, **kwargs)¶: Warning

This operator is now deprecated. Use readers.nemo_asr() instead.

In DALI 1.0 all readers were moved into a dedicated readers submodule and renamed to follow a common pattern. This is a placeholder operator with identical functionality to allow for backward compatibility.

Legacy alias for readers.nemo_asr().

nvidia.dali.fn.nonsilent_region(*inputs, **kwargs)¶

Performs leading and trailing silence detection in an audio buffer.

The operator returns the beginning and length of the non-silent region by comparing the short term power calculated for window_length of the signal with a silence cut-off threshold. The signal is considered to be silent when the short_term_power_db is less than the cutoff_db. where:

short_term_power_db = 10 * log10( short_term_power / reference_power )

Unless specified otherwise, reference_power is the maximum power of the signal.

Inputs and outputs:

Input 0 - 1D audio buffer.
Output 0 - Index of the first sample in the nonsilent region.
Output 1 - Length of nonsilent region.

Note

If Outputs[1] == 0, the value in Outputs[0] is undefined.

Supported backends

‘cpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
cutoff_db (float, optional, default = -60.0) – The threshold, in dB, below which the signal is considered silent.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
reference_power (float, optional, default = 0.0) –
The reference power that is used to convert the signal to dB.

If a value is not provided, the maximum power of the signal will be used as the reference.
reset_interval (int, optional, default = 8192) –
The number of samples after which the moving mean average is recalculated to avoid loss of precision.

If reset_interval == -1, or the input type allows exact calculation, the average will not be reset. The default value can be used for most of the use cases.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
window_length (int, optional, default = 2048) – Size of the sliding window used to calculate of the short-term power of the signal.

nvidia.dali.fn.normal_distribution(*inputs, **kwargs)¶

Warning

This operator is now deprecated. Use random.normal() instead.

Generates random numbers following a normal distribution.

The shape of the generated data can be either specified explicitly with a shape argument, or chosen to match the shape of the input, if provided. If none are present, a single value per sample is generated.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList, optional) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional) –
Output data type.

Note

The generated numbers are converted to the output data type, rounding and clamping if necessary.
mean (float or TensorList of float, optional, default = 0.0) – Mean of the distribution.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shape (int or list of int or TensorList of int, optional) – Shape of the output data.
stddev (float or TensorList of float, optional, default = 1.0) – Standard deviation of the distribution.

nvidia.dali.fn.normalize(*inputs, **kwargs)¶

Normalizes the input by removing the mean and dividing by the standard deviation.

The mean and standard deviation can be calculated internally for the specified subset of axes or can be externally provided as the mean and stddev arguments.

The normalization is done following the formula:

out = scale * (in - mean) / stddev + shift

The formula assumes that out and in are equally shaped tensors, but mean and stddev might be either tensors of same shape, scalars, or a mix of these.

Note

The expression follows the numpy broadcasting rules.

Sizes of the non-scalar mean and stddev must have an extent of 1, if given axis is reduced, or match the corresponding extent of the input. A dimension is considered reduced if it is listed in axes or axis_names. If neither the axes nor the axis_names argument is present, the set of reduced axes is inferred by comparing the input shape to the shape of the mean/stddev arguments, but the set of reduced axes must be the same for all tensors in the batch.

Here are some examples of valid argument combinations:

Per-sample normalization of dimensions 0 and 2:

axes = 0,2                                        # optional
input.shape = [ [480, 640, 3], [1080, 1920, 4] ]
batch = False
mean.shape =  [ [1, 640, 1], [1, 1920, 1] ]
stddev = (not supplied)

With these shapes, batch normalization is not possible, because the non-reduced dimension has a different extent across samples.

Batch normalization of dimensions 0 and 1:

axes = 0,1                                        # optional
input.shape = [ [480, 640, 3], [1080, 1920, 3] ]
batch = True
mean = (scalar)
stddev.shape =  [ [1, 1, 3] ] ]

For color images, this example normalizes the 3 color channels separately, but across all samples in the batch.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

axes (int or list of int, optional, default = []) –
Indices of dimensions along which the input is normalized.

By default, all axes are used, and the axes can also be specified by name. See axis_names for more informaton.
axis_names (layout str, optional, default = ‘’) –
Names of the axes in the input.

Axis indices are taken from the input layout, and this argument cannot be used with axes.
batch (bool, optional, default = False) –
If set to True, the mean and standard deviation are calculated across tensors in the batch.

This argument also requires that the input sample shapes in the non-reduced axes match.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
ddof (int, optional, default = 0) –
Delta Degrees of Freedom for Bessel’s correction.

The variance is estimated by using the following formula:
```
sum(Xi - mean)**2 / (N - ddof).
```
This argument is ignored when an externally supplied standard deviation is used.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.FLOAT) –
Output data type.

When using integral types, use shift and scale to improve the usage of the output type’s dynamic range. If dtype is an integral type, out of range values are clamped, and non-integer values are rounded to nearest integer.
epsilon (float, optional, default = 0.0) – A value that is added to the variance to avoid division by small numbers.
mean (float or TensorList of float, optional) –
Mean value to be subtracted from the data.

The value can be a scalar or a batch of tensors with the same dimensionality as the input. The extent in each dimension must match the value of the input or be equal to 1. If the extent is 1, the value will be broadcast in this dimension. If the value is not specified, the mean is calculated from the input. A non-scalar mean cannot be used when batch argument is set to True.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
scale (float, optional, default = 1.0) –
The scaling factor applied to the output.

This argument is useful for integral output types.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shift (float, optional, default = 0.0) –
The value to which the mean will map in the output.

This argument is useful for unsigned output types.
stddev (float or TensorList of float, optional) –
Standard deviation value to scale the data.

See mean argument for more information about shape constraints. If a value is not specified, the standard deviation is calculated from the input. A non-scalar stddev cannot be used when batch argument is set to True.

nvidia.dali.fn.numpy_reader(*inputs, **kwargs)¶: Warning

This operator is now deprecated. Use readers.numpy() instead.

In DALI 1.0 all readers were moved into a dedicated readers submodule and renamed to follow a common pattern. This is a placeholder operator with identical functionality to allow for backward compatibility.

Legacy alias for readers.numpy().

nvidia.dali.fn.one_hot(*inputs, **kwargs)¶

Produces a one-hot encoding of the input.

Adds a new axis or converts scalar input into an axis of num_classes elements.

For given input coordinate (x0, x1, ..., xn), and axis = k, the output sample is specified as:

cls = input[x0, x1, ..., xn]
output[x0, x1, ..., xk-1, i, xk, ..., xn] = on_value if i == cls else off_value

for all i in range [0, num_classes).

For scalars, the output is set to on_value at the index taken from input and off_value elsewhere:

output[i] = on_value if i == input else off_value

For backward compatibility, any input in which all tensors have only one element (regardless of the number of dimensions) is considered scalar. Legacy interpretation of tensors as scalars is not supported if axis argument is specified.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

axis (int, optional, default = -1) – Dimension to place the one-hot encoding axis of num_classes size. By default it’s appended as the last dimension for non-scalar inputs. For scalar inputs, it becomes the only dimension.
axis_name (str, optional) – Single character that will be used as a name for the newly added dimension in the output layout. If no character is provided, the output layout will be empty.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.FLOAT) – Output data type.
num_classes (int, optional, default = 0) – Number of all classes in the data.
off_value (float, optional, default = 0.0) –
Value that will be used to fill the output to indicate the lack of given class in the corresponding input coordinate.

This value will be cast to the dtype type.
on_value (float, optional, default = 1.0) –
Value that will be used to fill the output to indicate given class in the corresponding input coordinate.

This value will be cast to the dtype type.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.optical_flow(*inputs, **kwargs)¶

Calculates the optical flow between images in the input.

The main input for this operator is a sequence of frames. Optionally, the operator can be provided with external hints for the optical flow calculation. The output format of this operator matches the output format of the optical flow driver API. Refer to https://developer.nvidia.com/opticalflow-sdk for more information about the Turing and Ampere optical flow hardware that is used by DALI.

This operator allows sequence inputs.

Supported backends

‘gpu’

Parameters

input0 (TensorList ('FHWC')) – Input to the operator.
input1 (TensorList ('FHWC'), optional) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
enable_external_hints (bool, optional, default = False) –
Enables or disables the external hints for optical flow calculations.

External hints are analogous to temporal hints, but the only difference is that external hints come from an external source. When this option is enabled, the operator requires two inputs.
enable_temporal_hints (bool, optional, default = False) –
Enables or disables temporal hints for sequences that are longer than two images.

The hints are used to improve the quality of the output motion field as well as to speed up the calculations. The hints are especially useful in presence of large displacements or periodic patterns which might confuse the optical flow algorithms. )
hint_grid (int, optional, default = 4) –
Sets the grid size for the hint vector field.

The hints are used to improve the quality of the output motion field as well as to speed up the calculations. The grid resolution could be set to a different value than the output.

Note

Currently, only a 1, 2, 4 and 8 are supported for Ampere and 4 for Turing.
image_type (nvidia.dali.types.DALIImageType, optional, default = DALIImageType.RGB) – Input color space (RGB, BGR or GRAY).
output_grid (int, optional, default = 4) –
Sets the grid size for the output vector field.

This operator produces the motion vector field at a coarser resolution than the input pixels. This parameter specifies the size of the pixel grid cell corresponding to one motion vector. For example, a value of 4 will produce one motion vector for each 4x4 pixel block.

Note

Currently, only a 1, 2 and 4 are supported for Ampere and 4 for Turing.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
preset (float, optional, default = 0.0) –
Speed and quality level of the optical flow calculation.

Allowed values are:
- 0.0 is the lowest speed and the best quality.
- 0.5 is the medium speed and quality.
- 1.0 is the fastest speed and the lowest quality.
The lower the speed, the more additional pre- and postprocessing is used to enhance the quality of the optical flow result.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
output_format (int) –

Warning

The argument output_format is a deprecated alias for output_grid. Use output_grid instead.

nvidia.dali.fn.pad(*inputs, **kwargs)¶

Pads all samples with the fill_value in the specified axes to match the biggest extent in the batch for those axes or to match the minimum shape specified.

Here are a few examples:

1-D samples, fill_value = -1, axes = (0,)

The samples are padded in the first axis to match the extent of the largest sample.

input  = [[3,   4,   2,   5,   4],
          [2,   2],
          [3, 199,   5]];
output = [[3,   4,   2,   5,   4],
          [2,   2,  -1,  -1,  -1],
          [3, 199,   5,  -1,  -1]]

1-D samples, fill_value = -1, axes = (0,), shape = (7,)

The samples are padded in the first axis to a minimum extent of 7.

input  = [[3,   4,   2,   5,   4],
          [2,   2],
          [3, 199,   5],
          [1,   2,   3,   4,   5,   6,   7,   8]];
output = [[3,   4,   2,   5,   4,  -1,  -1],
          [2,   2,  -1,  -1,  -1,  -1,  -1],
          [3, 199,   5,  -1,  -1,  -1,  -1],
          [1,   2,   3,   4,   5,   6,   7,   8]]

1-D samples, fill_value = -1, axes = (0,), align = (4,)

The samples are padded in the first axis to match the extent of the largest sample and the alignment requirements. The output extent is 8, which is a result of rounding up the largest extent (5) to a multiple of alignment (4).

input  = [[3,   4,   2,   5,   4],
          [2,   2],
          [3, 199,   5]];
output = [[3,   4,   2,   5,   4,  -1,  -1,  -1],
          [2,   2,  -1,  -1,  -1,  -1,  -1,  -1],
          [3, 199,   5,  -1,  -1,  -1,  -1,  -1]]

1-D samples, fill_value = -1, axes = (0,), shape = (1,), align = (2,)

The samples are padded in the first axis to match the alignments requirements only. The minimum extent (shape) is set to 1 to avoid any padding other than the necessary for alignment.

input  = [[3,   4,   2,   5,   4],
          [2,   2],
          [3, 199,   5]];
output = [[3,   4,   2,   5,   4,  -1],
          [2,   2],
          [3, 199,   5,  -1]]

2-D samples, fill_value = 42, axes = (1,)

The samples are padded in the second axis to match the extent of the largest sample and uses a custom fill value 42 instead of the default 0.

input  = [[[1,  2,  3,  4],
           [5,  6,  7,  8]],
          [[1,  2],
           [4,  5]]]
output = [[[1,  2,  3,  4],
           [5,  6,  7,  8]],
          [[1,  2, 42, 42],
           [4,  5, 42, 42]]]

2-D samples, fill_value = 0, axes = (0, 1), align = (4, 5)

The samples are padded in the first and second axes to match the alignment requirements of each axis.

input  = [[[1,  2,  3,  4],
           [5,  6,  7,  8],
           [9, 10, 11, 12]],
          [[1, 2],
           [4, 5]]]
output = [[[1,  2,  3,  4,  0],
           [5,  6,  7,  8,  0],
           [9, 10, 11, 12,  0],
           [0,  0,  0,  0,  0]],
          [[1,  2,  0,  0,  0],
           [4,  5,  0,  0,  0],
           [0,  0,  0,  0,  0],
           [0,  0,  0,  0,  0]]]

2-D samples, fill_value = 0, axes = (0, 1), align = (1, 2), shape = (4, -1)

The samples are padded in the first axis to match a minimum extent of 4, and in the second axis to match the largest sample in the batch and an alignment of 2.

input  = [[[1,  2,  3],
           [4,  5,  6]],
          [[1, 2],
           [4, 5],
           [6, 7]]]
output = [[[1,  2,  3,  0],
           [4,  5,  6,  0],
           [0,  0,  0,  0],
           [0,  0,  0,  0]],
          [[1,  2,  0,  0],
           [4,  5,  0,  0],
           [6,  7,  0,  0],
           [0,  0,  0,  0]]]

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

align (int or list of int or TensorList of int, optional, default = []) –
If specified, this argument determines the alignment on the dimensions specified by axes or axis_names.

The extent on axis = axes[i] will be adjusted to be a multiple of align[i].

If an integer value is provided, the alignment restrictions are applied to all the padded axes.

To use alignment only, that is without any default or explicit padding behavior, set the minimum shape to 1 for the specified axis.
axes (int or list of int or TensorList of int, optional, default = []) –
Indices of the axes on which the batch samples will be padded.

Negative values are interpreted as counting dimensions from the back. Valid range: [-ndim, ndim-1], where ndim is the number of dimensions in the input data.

The axis_names and axes arguments are mutually exclusive. If axes and axis_names are empty, or have not been provided, the output will be padded on all of the axes.
axis_names (layout str, optional, default = ‘’) –
Names of the axes on which the batch samples will be padded.

The axis_names and axes arguments are mutually exclusive. If axes and axis_names are empty, or have not been provided, the output will be padded on all of the axes.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
fill_value (float or TensorList of float, optional, default = 0.0) – The value to pad the batch with.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shape (int or list of int or TensorList of int, optional, default = []) –
The extents of the output shape in the axes specified by the axes or axis_names.

Specifying -1 for an axis restores the default behavior of extending the axis to accommodate the aligned size of the largest sample in the batch.

If the provided extent is smaller than the one of the samples, padding will be applied only to match the required alignment. For example, to disable padding in an axis, except for the necessary for alignment, you can specify a value of 1.

nvidia.dali.fn.paste(*inputs, **kwargs)¶

Pastes the input images on a larger canvas, where the canvas size is equal to input size * ratio.

Supported backends

‘gpu’

Parameters

input (TensorList ('HWC')) – Input to the operator.

Keyword Arguments

fill_value (int or list of int) –
Tuple of the values of the color that is used to fill the canvas.

The length of the tuple must be equal to n_channels.
ratio (float or TensorList of float) – Ratio of canvas size to input size. Must be >= 1.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
min_canvas_size (float or TensorList of float, optional, default = 0.0) – Enforces the minimum paste canvas dimension after scaling the input size by the ratio.
n_channels (int, optional, default = 3) – Number of channels in the image.
paste_x (float or TensorList of float, optional, default = 0.5) – Horizontal position of the paste in (0.0 - 1.0) image coordinates.
paste_y (float or TensorList of float, optional, default = 0.5) – Vertical position of the paste in (0.0 - 1.0) image coordinates.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.peek_image_shape(*inputs, **kwargs)¶

Obtains the shape of the encoded image.

Supported backends

‘cpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
type (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.INT64) – Data type, to which the sizes are converted.

nvidia.dali.fn.permute_batch(*inputs, **kwargs)¶

Returns a batch of tensors constructed by selecting tensors from the input based on indices given in indices argument:

out_tensor[i] = in_tensor[indices[i]]

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

indices (int or list of int or TensorList of int) –
List of indices, matching current batch size, or a batch of scalars representing indices of the tensors in the input batch.

The indices must be within [0..batch_size) range. Repetitions and omissions are allowed.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.power_spectrum(*inputs, **kwargs)¶

Calculates power spectrum of the signal.

Supported backends

‘cpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

axis (int, optional, default = -1) –
Index of the dimension to be transformed to the frequency domain.

By default, the last dimension is selected.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
nfft (int, optional) –
Size of the FFT.

By default, the nfft is selected to match the length of the data in the transformation axis.

The number of bins that are created in the output is calculated with the following formula:
```
nfft // 2 + 1
```
Note

The output only represents the positive part of the spectrum.
power (int, optional, default = 2) –
Exponent of the FFT magnitude.

The supported values are:
- 2 for power spectrum (real*real + imag*imag)
- 1 for the complex magnitude (sqrt(real*real + imag*imag)).
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.preemphasis_filter(*inputs, **kwargs)¶

Applies preemphasis filter to the input data.

This filter, in simple form, can be expressed by the formula:

Y[t] = X[t] - coeff * X[t-1]    if t > 1
Y[t] = X[t] - coeff * X_border  if t == 0

with X and Y being the input and output signal, respectively.

The value of X_border depends on the border argument:

X_border = 0                    if border_type == 'zero'
X_border = X[0]                 if border_type == 'clamp'
X_border = X[1]                 if border_type == 'reflect'

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

border (str, optional, default = ‘clamp’) – Border value policy. Possible values are "zero", "clamp", "reflect".
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.FLOAT) – Data type for the output.
preemph_coeff (float or TensorList of float, optional, default = 0.97) – Preemphasis coefficient coeff.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.python_function(*inputs, **kwargs)¶

Executes a Python function.

This operator can be used to execute custom Python code in the DALI pipeline. The function receives the data from DALI as NumPy arrays in case of CPU operators or as CuPy arrays for GPU operators. It is expected to return the results in the same format. For a more universal data format, see nvidia.dali.fn.dl_tensor_python_function(). The function should not modify input tensors.

Warning

Currently, this operator can be used only in pipelines with the exec_async=False and exec_pipelined=False values specified and should only be used for prototyping and debugging.

Warning

This operator is not compatible with TensorFlow integration.

This operator allows sequence inputs and supports volumetric data.

This operator will not be optimized out of the graph.

Supported backends

‘cpu’
‘gpu’

Parameters

input[0..255] (TensorList, optional) – This function accepts up to 256 optional positional inputs

Keyword Arguments

function (object) – Function object.
batch_processing (bool, optional, default = False) –
Determines whether the function is invoked once per batch or separately for every sample in the batch.

If set to True, the function will receive its arguments as lists of NumPy or CuPy arrays, for CPU and GPU backend, respectively.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
num_outputs (int, optional, default = 1) – Number of outputs.
output_layouts (layout str or list of layout str, optional) –
Tensor data layouts for the outputs.

This argument can be a list that contains a distinct layout for each output. If the list has fewer than num_outputs elements, only the first outputs have the layout set and the rest of the outputs have no layout assigned.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.random_bbox_crop(*inputs, **kwargs)¶

Applies a prospective random crop to an image coordinate space while keeping the bounding boxes, and optionally labels, consistent.

This means that after applying the random crop operator to the image coordinate space, the bounding boxes will be adjusted or filtered out to match the cropped ROI. The applied random crop operation is constrained by the arguments that are provided to the operator.

The cropping window candidates are randomly selected until one matches the overlap restrictions that are specified by the thresholds argument. thresholds values represent a minimum overlap metric that is specified by threshold_type, such as the intersection-over-union of the cropping window and the bounding boxes or the relative overlap as a ratio of the intersection area and the bounding box area.

Additionally, if allow_no_crop is True, the cropping may be skipped entirely as one of the valid results of the operator.

The following modes of a random crop are available:

Randomly shaped window, which is randomly placed in the original input space.

The random crop window dimensions are selected to satisfy the aspect ratio and relative area restrictions.

If input_shape is provided, it will be taken into account for the aspect ratio range check.

Otherwise, the aspect ratios are calculated in relative terms.

In other words, without input_shape, an aspect ratio of 1.0 is equivalent to the aspect ratio of the input image.
Fixed size window, which is randomly placed in the original input space.

The random crop window dimensions are taken from the crop_shape argument and the anchor is

randomly selected.

When providing crop_shape, input_shape is also required (these dimensions are required to

scale the output bounding boxes).

The num_attempts argument can be used to control the maximum number of attempts to produce a valid crop to match a minimum overlap metric value from thresholds.

Warning

When allow_no_crop is False and thresholds does not contain 0.0, if you do not increase the num_attempts value, it might continue to loop for a long time.

Inputs: 0: bboxes, (1: labels)

The first input, bboxes, refers to the bounding boxes that are provided as a two-dimensional tensor where the first dimension refers to the index of the bounding box, and the second dimension refers to the index of the coordinate.

The coordinates are relative to the original image dimensions (that means, a range of [0.0, 1.0]) that represent the start and, depending on the value of bbox_layout, the end of the region or start and shape. For example, bbox_layout=”xyXY” means the bounding box coordinates follow the start_x, start_y, end_x, and end_y order, and bbox_layout=”xyWH” indicates that the order is start_x, start_y, width, and height. See the bbox_layout argument description for more information.

An optional input labels can be provided, representing the labels that are associated with each of the bounding boxes.

Outputs: 0: anchor, 1: shape, 2: bboxes (, 3: labels, 4: bboxes_indices)

The resulting crop parameters are provided as two separate outputs, anchor and shape, that can be fed directly to the nvidia.dali.fn.slice() operator to complete the cropping of the original image. anchor and shape contain the starting coordinates and dimensions for the crop in the [x, y, (z)] and [w, h, (d)] formats, respectively. The coordinates can be represented in absolute or relative terms, and the representation depends on whether the fixed crop_shape was used.

Note

Both anchor and shape are returned as a float, even if they represent absolute coordinates due to providing crop_shape argument. In order for them to be interpreted correctly by nvidia.dali.fn.slice(), normalized_anchor and normalized_shape should be set to False.

The third output contains the bounding boxes, after filtering out the ones with a centroid outside of the cropping window, and with the coordinates mapped to the new coordinate space.

The next output is optional, and it represents the labels associated with the filtered bounding boxes. The output will be present if a labels input was provided.

The last output, also optional, correspond to the original indices of the bounding boxes that passed the centroid filter and are present in the output. This output will be present if the option output_bbox_indices is set to True.

Supported backends

‘cpu’

Parameters

boxes (2D TensorList of float) – Relative coordinates of the bounding boxes that are represented as a 2D tensor, where the first dimension refers to the index of the bounding box, and the second dimension refers to the index of the coordinate.
labels (1D TensorList of integers, optional) – Labels that are associated with each of the bounding boxes.

Keyword Arguments

all_boxes_above_threshold (bool, optional, default = True) –
If set to True, all bounding boxes in a sample should overlap with the cropping window as specified by thresholds.

If the bounding boxes do not overlap, the cropping window is considered to be invalid. If set to False, and at least one bounding box overlaps the window, the window is considered to be valid.
allow_no_crop (bool, optional, default = True) – If set to True, one of the possible outcomes of the random process will be to not crop, as if the outcome was one more thresholds value from which to choose.
aspect_ratio (float or list of float, optional, default = [1.0, 1.0]) –
Valid range of aspect ratio of the cropping windows.

This parameter can be specified as either two values (min, max) or six values (three pairs), depending on the dimensionality of the input.
- For 2D bounding boxes, one range of valid aspect ratios (x/y) should be provided (e.g. [min_xy, max_xy]).
- For 3D bounding boxes, three separate aspect ratio ranges may be specified, for x/y, x/z and y/z pairs of dimensions.
  
  They are provided in the following order [min_xy, max_xy, min_xz, max_xz, min_yz, max_yz]. Alternatively, if only one aspect ratio range is provided, it will be used for all three pairs of dimensions.
The value for min should be greater than 0.0, and min should be less than or equal to the max value. By default, square windows are generated.

Note

Providing aspect_ratio and scaling is incompatible with explicitly specifying crop_shape.

Note

If input_shape is provided, it will be taken into account for the calculation of the cropping window aspect ratio. Otherwise, the aspect ratio ranges are relative to the image dimensions. In other words, when input_shape is not specified, an aspect ratio of 1.0 is equivalent to the original aspect ratio of the image.
bbox_layout (layout str, optional, default = ‘’) –
Determines the meaning of the coordinates of the bounding boxes.

The value of this argument is a string containing the following characters:
```
x (horizontal start anchor), y (vertical start anchor), z (depthwise start anchor),
X (horizontal end anchor),   Y (vertical end anchor),   Z (depthwise end anchor),
W (width),                   H (height),                D (depth).
```
Note

If this value is left empty, depending on the number of dimensions, “xyXY” or “xyzXYZ” is assumed.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
crop_shape (int or list of int or TensorList of int, optional, default = []) –
If provided, the random crop window dimensions will be fixed to this shape.

The order of dimensions is determined by the layout provided in shape_layout.

Note

When providing crop_shape, input_shape should be provided as well. Providing explicit crop_shape is incompatible with using scaling and aspect_ratio arguments.
input_shape (int or list of int or TensorList of int, optional, default = []) –
Specifies the shape of the original input image.

The order of dimensions is determined by the layout that is provided in shape_layout.
ltrb (bool, optional, default = True) –
If set to True, bboxes are returned as [left, top, right, bottom]; otherwise they are provided as [left, top, width, height].

Warning

This argument has been deprecated. To specify the bbox encoding, use bbox_layout instead. For example, ltrb=True is equal to bbox_layout=”xyXY”, and ltrb=False corresponds to bbox_layout=”xyWH”.
num_attempts (int, optional, default = 1) –
Number of attempts to get a crop window that matches the aspect_ratio and a selected value from thresholds.

After each num_attempts, a different threshold will be picked, until the threshold reaches a maximum of total_num_attempts (if provided) or otherwise indefinitely.
output_bbox_indices (bool, optional, default = False) – If set to True, an extra output will be returned, containing the original indices of the bounding boxes that passed the centroid filter and are present in the output bounding boxes.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
scaling (float or list of float, optional, default = [1.0, 1.0]) –
Range [min, max] for the crop size with respect to the original image dimensions.

The value of min and max must satisfy the condition 0.0 <= min <= max.

Note

Providing aspect_ratio and scaling is incompatible when explicitly specifying the crop_shape value.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shape_layout (layout str, optional, default = ‘’) –
Determines the meaning of the dimensions provided in crop_shape and input_shape.

The values are:
- W (width)
- H (height)
- D (depth)
Note

If left empty, depending on the number of dimensions "WH" or "WHD" will be assumed.
threshold_type (str, optional, default = ‘iou’) –
Determines the meaning of thresholds.

By default, thresholds refers to the intersection-over-union (IoU) of the bounding boxes with respect to the cropping window. Alternatively, the threshold can be set to “overlap” to specify the fraction (by area) of the bounding box that will will fall inside the crop window. For example, a threshold value of 1.0 means the entire bounding box must be contained in the resulting cropping window.
thresholds (float or list of float, optional, default = [0.0]) –
Minimum IoU or a different metric, if specified by threshold_type, of the bounding boxes with respect to the cropping window.

Each sample randomly selects one of the thresholds, and the operator will complete up to the specified number of attempts to produce a random crop window that has the selected metric above that threshold. See num_attempts for more information about configuring the number of attempts.
total_num_attempts (int, optional, default = -1) –
If provided, it indicates the total maximum number of attempts to get a crop window that matches the aspect_ratio and any selected value from thresholds.

After total_num_attempts attempts, the best candidate will be selected.

If this value is not specified, the crop search will continue indefinitely until a valid crop is found.

Warning

If you do not provide a total_num_attempts value, this can result in an infinite loop if the conditions imposed by the arguments cannot be satisfied.

nvidia.dali.fn.random_resized_crop(*inputs, **kwargs)¶

Performs a crop with a randomly selected area and aspect ratio and resizes it to the specified size.

Expects a three-dimensional input with samples in height, width, channels (HWC) layout.

This operator allows sequence inputs.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList ('HWC', 'CHW', 'FHWC', 'FCHW', 'CFHW')) – Input to the operator.

Keyword Arguments

size (int or list of int) – Size of the resized image.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional) –
Output data type.

Must be same as input type or float. If not set, input type is used.
interp_type (nvidia.dali.types.DALIInterpType or TensorList of nvidia.dali.types.DALIInterpType, optional, default = DALIInterpType.INTERP_LINEAR) –
Type of interpolation to be used.

Use min_filter and mag_filter to specify different filtering for downscaling and upscaling.
mag_filter (nvidia.dali.types.DALIInterpType or TensorList of nvidia.dali.types.DALIInterpType, optional, default = DALIInterpType.INTERP_LINEAR) – Filter used when scaling up.
min_filter (nvidia.dali.types.DALIInterpType or TensorList of nvidia.dali.types.DALIInterpType, optional, default = DALIInterpType.INTERP_LINEAR) – Filter used when scaling down.
minibatch_size (int, optional, default = 32) – Maximum number of images that are processed in a kernel call.
num_attempts (int, optional, default = 10) – Maximum number of attempts used to choose random area and aspect ratio.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
random_area (float or list of float, optional, default = [0.08, 1.0]) –
Range from which to choose random area fraction A.

The cropped image’s area will be equal to A * original image’s area.
random_aspect_ratio (float or list of float, optional, default = [0.75, 1.333333]) – Range from which to choose random aspect ratio (width/height).
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
temp_buffer_hint (int, optional, default = 0) –
Initial size in bytes, of a temporary buffer for resampling.

Note

This argument is ignored for the CPU variant.

nvidia.dali.fn.reinterpret(*inputs, **kwargs)¶

Treats content of the input as if it had a different type, shape, and/or layout.

The buffer contents are not copied.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

data (TensorList) – Data to be reshaped
shape_input (1D TensorList of integers, optional) – Same as shape keyword argument

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.NO_TYPE) –
Output data type.

The total size, in bytes, of the output must match the input. If no shape is provided, the innermost dimension is adjusted accordingly. If the byte size of the innermost dimension is not divisible by the size of the target type, an error occurs.
layout (layout str, optional, default = ‘’) –
New layout for the data.

If a value is not specified, if number of dimension matches existing layout, the output layout is preserved. If the number of dimensions does not match, the argument is reset to empty. If a value is set, and is not empty, the layout must match the dimensionality of the output.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
rel_shape (float or list of float or TensorList of float, optional, default = []) –
The relative shape of the output.

Number of dimensions cannot exceed the number of dimensions of the input. There can be one negative extent that receives the size that is required to match the input volume. For example, an input of shape [480, 640, 3] and a rel_shape = [0.5, -1] results in the shape [240, 3840].

Note

rel_shape and shape are mutually exclusive.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shape (int or list of int or TensorList of int, optional, default = []) –
The desired shape of the output.

Number of dimensions cannot exceed the number of dimensions of the input. There can be one negative extent that receives the size that is required to match the input volume. For example, an input of shape [480, 640, 3] and shape = [240, -1] results in the shape [240, 3840].

Note

rel_shape and shape are mutually exclusive.
src_dims (int or list of int or TensorList of int, optional, default = []) –
Indices of dimensions to keep.

This argument can be used to manipulate the order of existing dimensions or to remove or add dimension. A special index value -1 can be used to insert new dimensions.

For example, reshaping a sample with shape [300, 200, 1] and a src_dims argument [-1, 1, 0] produces an output shape [1, 200, 300]. A leading dimension with extent 1 is inserted at the beginning, followed by the first original dimensions but in reverse order. The last dimension is removed.

The src_dims argument can be used together with rel_shape, in which case the relative extents in rel_shape describe to the target dimensions. In the example above, specifying rel_shape = [-1, 0.5, 2] would result in the output shape [1, 100, 600].

All indices must be in the range of valid dimensions of the input, or -1.

nvidia.dali.fn.reshape(*inputs, **kwargs)¶

Treats content of the input as if it had a different shape and/or layout.

The buffer contents are not copied.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

data (TensorList) – Data to be reshaped
shape_input (1D TensorList of integers, optional) – Same as shape keyword argument

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
layout (layout str, optional, default = ‘’) –
New layout for the data.

If a value is not specified, if number of dimension matches existing layout, the output layout is preserved. If the number of dimensions does not match, the argument is reset to empty. If a value is set, and is not empty, the layout must match the dimensionality of the output.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
rel_shape (float or list of float or TensorList of float, optional, default = []) –
The relative shape of the output.

Number of dimensions cannot exceed the number of dimensions of the input. There can be one negative extent that receives the size that is required to match the input volume. For example, an input of shape [480, 640, 3] and a rel_shape = [0.5, -1] results in the shape [240, 3840].

Note

rel_shape and shape are mutually exclusive.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shape (int or list of int or TensorList of int, optional, default = []) –
The desired shape of the output.

Number of dimensions cannot exceed the number of dimensions of the input. There can be one negative extent that receives the size that is required to match the input volume. For example, an input of shape [480, 640, 3] and shape = [240, -1] results in the shape [240, 3840].

Note

rel_shape and shape are mutually exclusive.
src_dims (int or list of int or TensorList of int, optional, default = []) –
Indices of dimensions to keep.

This argument can be used to manipulate the order of existing dimensions or to remove or add dimension. A special index value -1 can be used to insert new dimensions.

For example, reshaping a sample with shape [300, 200, 1] and a src_dims argument [-1, 1, 0] produces an output shape [1, 200, 300]. A leading dimension with extent 1 is inserted at the beginning, followed by the first original dimensions but in reverse order. The last dimension is removed.

The src_dims argument can be used together with rel_shape, in which case the relative extents in rel_shape describe to the target dimensions. In the example above, specifying rel_shape = [-1, 0.5, 2] would result in the output shape [1, 100, 600].

All indices must be in the range of valid dimensions of the input, or -1.

nvidia.dali.fn.resize(*inputs, **kwargs)¶

Resize images.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList ('HWC', 'FHWC', 'CHW', 'FCHW', 'CFHW', 'DHWC', 'FDHWC', 'CDHW', 'FCDHW', 'CFDHW')) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional) –
Output data type.

Must be same as input type or float. If not set, input type is used.
image_type (nvidia.dali.types.DALIImageType) –

Warning

The argument image_type is no longer used and will be removed in a future release.
interp_type (nvidia.dali.types.DALIInterpType or TensorList of nvidia.dali.types.DALIInterpType, optional, default = DALIInterpType.INTERP_LINEAR) –
Type of interpolation to be used.

Use min_filter and mag_filter to specify different filtering for downscaling and upscaling.
mag_filter (nvidia.dali.types.DALIInterpType or TensorList of nvidia.dali.types.DALIInterpType, optional, default = DALIInterpType.INTERP_LINEAR) – Filter used when scaling up.
max_size (float or list of float, optional) –
Limit of the output size.

When the operator is configured to keep aspect ratio and only the smaller dimension is specified, the other(s) can grow very large. This can happen when using resize_shorter argument or “not_smaller” mode or when some extents are left unspecified.

This parameter puts a limit to how big the output can become. This value can be specified per-axis or uniformly for all axes.

Note

When used with “not_smaller” mode or resize_shorter argument, max_size takes precedence and the aspect ratio is kept - for example, resizing with mode="not_smaller", size=800, max_size=1400 an image of size 1200x600 would be resized to 1400x700.
min_filter (nvidia.dali.types.DALIInterpType or TensorList of nvidia.dali.types.DALIInterpType, optional, default = DALIInterpType.INTERP_LINEAR) – Filter used when scaling down.
minibatch_size (int, optional, default = 32) – Maximum number of images that are processed in a kernel call.
mode (str, optional, default = ‘default’) –
Resize mode.

Here is a list of supported modes:
- "default" - image is resized to the specified size.
  
  Missing extents are scaled with the average scale of the provided ones.
- "stretch" - image is resized to the specified size.
  
  Missing extents are not scaled at all.
- "not_larger" - image is resized, keeping the aspect ratio, so that no extent of the output image exceeds the specified size.
  
  For example, a 1280x720, with a desired output size of 640x480, actually produces a 640x360 output.
- "not_smaller" - image is resized, keeping the aspect ratio, so that no extent of the output image is smaller than specified.
  
  For example, a 640x480 image with a desired output size of 1920x1080, actually produces a 1920x1440 output.
  
  This argument is mutually exclusive with resize_longer and resize_shorter
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
resize_longer (float or TensorList of float, optional, default = 0.0) –
The length of the longer dimension of the resized image.

This option is mutually exclusive with resize_shorter and explicit size arguments, and the operator keeps the aspect ratio of the original image. This option is equivalent to specifying the same size for all dimensions and mode="not_larger".
resize_shorter (float or TensorList of float, optional, default = 0.0) –
The length of the shorter dimension of the resized image.

This option is mutually exclusive with resize_longer and explicit size arguments, and the operator keeps the aspect ratio of the original image. This option is equivalent to specifying the same size for all dimensions and mode="not_smaller". The longer dimension can be bounded by setting the max_size argument. See max_size argument doc for more info.
resize_x (float or TensorList of float, optional, default = 0.0) –
The length of the X dimension of the resized image.

This option is mutually exclusive with resize_shorter, resize_longer and size. If the resize_y is unspecified or 0, the operator keeps the aspect ratio of the original image. A negative value flips the image.
resize_y (float or TensorList of float, optional, default = 0.0) –
The length of the Y dimension of the resized image.

This option is mutually exclusive with resize_shorter, resize_longer and size. If the resize_x is unspecified or 0, the operator keeps the aspect ratio of the original image. A negative value flips the image.
resize_z (float or TensorList of float, optional, default = 0.0) –
The length of the Z dimension of the resized volume.

This option is mutually exclusive with resize_shorter, resize_longer and size. If the resize_x and resize_y are left unspecified or 0, then the op will keep the aspect ratio of the original volume. Negative value flips the volume.
roi_end (float or list of float or TensorList of float, optional) –
End of the input region of interest (ROI).

Must be specified together with roi_start. The coordinates follow the tensor shape order, which is the same as size. The coordinates can be either absolute (in pixels, which is the default) or relative (0..1), depending on the value of relative_roi argument. If the ROI origin is greater than the ROI end in any dimension, the region is flipped in that dimension.
roi_relative (bool, optional, default = False) – If true, ROI coordinates are relative to the input size, where 0 denotes top/left and 1 denotes bottom/right
roi_start (float or list of float or TensorList of float, optional) –
Origin of the input region of interest (ROI).

Must be specified together with roi_end. The coordinates follow the tensor shape order, which is the same as size. The coordinates can be either absolute (in pixels, which is the default) or relative (0..1), depending on the value of relative_roi argument. If the ROI origin is greater than the ROI end in any dimension, the region is flipped in that dimension.
save_attrs (bool, optional, default = False) – Save reshape attributes for testing.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
size (float or list of float or TensorList of float, optional) –
The desired output size.

Must be a list/tuple with one entry per spatial dimension, excluding video frames and channels. Dimensions with a 0 extent are treated as absent, and the output size will be calculated based on other extents and mode argument.
subpixel_scale (bool, optional, default = True) –
If True, fractional sizes, directly specified or calculated, will cause the input ROI to be adjusted to keep the scale factor.

Otherwise, the scale factor will be adjusted so that the source image maps to the rounded output size.
temp_buffer_hint (int, optional, default = 0) –
Initial size in bytes, of a temporary buffer for resampling.

Note

This argument is ignored for the CPU variant.

nvidia.dali.fn.resize_crop_mirror(*inputs, **kwargs)¶

Performs a fused resize, crop, mirror operation. Both fixed and random resizing and cropping are supported.

Supported backends

‘cpu’

Parameters

input (TensorList ('HWC')) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
crop (float or list of float, optional, default = [0.0, 0.0]) –
Shape of the cropped image, specified as a list of values (for example, (crop_H, crop_W) for the 2D crop and (crop_D, crop_H, crop_W) for the volumetric crop).

Providing crop argument is incompatible with providing separate arguments such as crop_d, crop_h, and crop_w.
crop_d (float or TensorList of float, optional, default = 0.0) –
Applies only to volumetric inputs; cropping window depth (in voxels).

crop_w, crop_h, and crop_d must be specified together. Providing values for crop_w, crop_h, and crop_d is incompatible with providing the fixed crop window dimensions (argument crop).
crop_h (float or TensorList of float, optional, default = 0.0) –
Cropping the window height (in pixels).

Providing values for crop_w and crop_h is incompatible with providing fixed crop window dimensions (argument crop).
crop_pos_x (float or TensorList of float, optional, default = 0.5) –
Normalized (0.0 - 1.0) horizontal position of the cropping window (upper left corner).

The actual position is calculated as crop_x = crop_x_norm * (W - crop_W), where crop_x_norm is the normalized position, W is the width of the image, and crop_W is the width of the cropping window.
crop_pos_y (float or TensorList of float, optional, default = 0.5) –
Normalized (0.0 - 1.0) vertical position of the start of the cropping window (typically, the upper left corner).

The actual position is calculated as crop_y = crop_y_norm * (H - crop_H), where crop_y_norm is the normalized position, H is the height of the image, and crop_H is the height of the cropping window.
crop_pos_z (float or TensorList of float, optional, default = 0.5) –
Applies only to volumetric inputs.

Normalized (0.0 - 1.0) normal position of the cropping window (front plane). The actual position is calculated as crop_z = crop_z_norm * (D - crop_D), where crop_z_norm is the normalized position, D is the depth of the image and crop_D is the depth of the cropping window.
crop_w (float or TensorList of float, optional, default = 0.0) –
Cropping window width (in pixels).

Providing values for crop_w and crop_h is incompatible with providing fixed crop window dimensions (argument crop).
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.NO_TYPE) –
Output data type.

Supported types: FLOAT, FLOAT16, and UINT8.

If not set, the input type is used.
fill_values (float or list of float, optional, default = [0.0]) –
Determines padding values and is only relevant if out_of_bounds_policy is set to “pad”.

If a scalar value is provided, it will be used for all the channels. If multiple values are provided, the number of values and channels must be identical (extent of dimension C in the layout) in the output slice.
image_type (nvidia.dali.types.DALIImageType) –

Warning

The argument image_type is no longer used and will be removed in a future release.
interp_type (nvidia.dali.types.DALIInterpType, optional, default = DALIInterpType.INTERP_LINEAR) – Type of interpolation used.
max_size (float or list of float, optional) –
Limit of the output size.

When the operator is configured to keep aspect ratio and only the smaller dimension is specified, the other(s) can grow very large. This can happen when using resize_shorter argument or “not_smaller” mode or when some extents are left unspecified.

This parameter puts a limit to how big the output can become. This value can be specified per-axis or uniformly for all axes.

Note

When used with “not_smaller” mode or resize_shorter argument, max_size takes precedence and the aspect ratio is kept - for example, resizing with mode="not_smaller", size=800, max_size=1400 an image of size 1200x600 would be resized to 1400x700.
mirror (int or TensorList of int, optional, default = 0) –
Mask for the horizontal flip.

Supported values:
- 0 - Do not perform horizontal flip for this image.
- 1 - Performs horizontal flip for this image.
mode (str, optional, default = ‘default’) –
Resize mode.

Here is a list of supported modes:
- "default" - image is resized to the specified size.
  
  Missing extents are scaled with the average scale of the provided ones.
- "stretch" - image is resized to the specified size.
  
  Missing extents are not scaled at all.
- "not_larger" - image is resized, keeping the aspect ratio, so that no extent of the output image exceeds the specified size.
  
  For example, a 1280x720, with a desired output size of 640x480, actually produces a 640x360 output.
- "not_smaller" - image is resized, keeping the aspect ratio, so that no extent of the output image is smaller than specified.
  
  For example, a 640x480 image with a desired output size of 1920x1080, actually produces a 1920x1440 output.
  
  This argument is mutually exclusive with resize_longer and resize_shorter
out_of_bounds_policy (str, optional, default = ‘error’) –
Determines the policy when slicing the out of bounds area of the input.

Here is a list of the supported values:
- "error" (default): Attempting to slice outside of the bounds of the input will produce an error.
- "pad": The input will be padded as needed with zeros or any other value that is specified with the fill_values argument.
- "trim_to_shape": The slice window will be cut to the bounds of the input.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
resize_longer (float or TensorList of float, optional, default = 0.0) –
The length of the longer dimension of the resized image.

This option is mutually exclusive with resize_shorter and explicit size arguments, and the operator keeps the aspect ratio of the original image. This option is equivalent to specifying the same size for all dimensions and mode="not_larger".
resize_shorter (float or TensorList of float, optional, default = 0.0) –
The length of the shorter dimension of the resized image.

This option is mutually exclusive with resize_longer and explicit size arguments, and the operator keeps the aspect ratio of the original image. This option is equivalent to specifying the same size for all dimensions and mode="not_smaller". The longer dimension can be bounded by setting the max_size argument. See max_size argument doc for more info.
resize_x (float or TensorList of float, optional, default = 0.0) –
The length of the X dimension of the resized image.

This option is mutually exclusive with resize_shorter, resize_longer and size. If the resize_y is unspecified or 0, the operator keeps the aspect ratio of the original image. A negative value flips the image.
resize_y (float or TensorList of float, optional, default = 0.0) –
The length of the Y dimension of the resized image.

This option is mutually exclusive with resize_shorter, resize_longer and size. If the resize_x is unspecified or 0, the operator keeps the aspect ratio of the original image. A negative value flips the image.
resize_z (float or TensorList of float, optional, default = 0.0) –
The length of the Z dimension of the resized volume.

This option is mutually exclusive with resize_shorter, resize_longer and size. If the resize_x and resize_y are left unspecified or 0, then the op will keep the aspect ratio of the original volume. Negative value flips the volume.
roi_end (float or list of float or TensorList of float, optional) –
End of the input region of interest (ROI).

Must be specified together with roi_start. The coordinates follow the tensor shape order, which is the same as size. The coordinates can be either absolute (in pixels, which is the default) or relative (0..1), depending on the value of relative_roi argument. If the ROI origin is greater than the ROI end in any dimension, the region is flipped in that dimension.
roi_relative (bool, optional, default = False) – If true, ROI coordinates are relative to the input size, where 0 denotes top/left and 1 denotes bottom/right
roi_start (float or list of float or TensorList of float, optional) –
Origin of the input region of interest (ROI).

Must be specified together with roi_end. The coordinates follow the tensor shape order, which is the same as size. The coordinates can be either absolute (in pixels, which is the default) or relative (0..1), depending on the value of relative_roi argument. If the ROI origin is greater than the ROI end in any dimension, the region is flipped in that dimension.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
size (float or list of float or TensorList of float, optional) –
The desired output size.

Must be a list/tuple with one entry per spatial dimension, excluding video frames and channels. Dimensions with a 0 extent are treated as absent, and the output size will be calculated based on other extents and mode argument.
subpixel_scale (bool, optional, default = True) –
If True, fractional sizes, directly specified or calculated, will cause the input ROI to be adjusted to keep the scale factor.

Otherwise, the scale factor will be adjusted so that the source image maps to the rounded output size.
output_dtype (nvidia.dali.types.DALIDataType) –

Warning

The argument output_dtype is a deprecated alias for dtype. Use dtype instead.

nvidia.dali.fn.roi_random_crop(*inputs, **kwargs)¶

Produces a fixed shape cropping window, randomly placed so that as much of the provided region of interest (ROI) is contained in it.

If the ROI is bigger than the cropping window, the cropping window will be a subwindow of the ROI. If the ROI is smaller than the cropping window, the whole ROI shall be contained in the cropping window.

If an input shape (in_shape) is given, the resulting cropping window is selected to be within the bounds of that input shape. Alternatively, the input data subject to cropping can be passed to the operator, in the operator. When providing an input shape, the region of interest should be within the bounds of the input and the cropping window shape should not be larger than the input shape.

If no input shape is provided, the resulting cropping window is unbounded, potentially resulting in out of bounds cropping.

The cropping window dimensions should be explicitly provided (crop_shape), and the ROI should be either specified with roi_start/roi_end or roi_start/roi_shape.

The operator produces an output representing the cropping window start coordinates.

Supported backends

‘cpu’

Parameters

input (TensorList, optional) – Input to the operator.

Keyword Arguments

crop_shape (int or list of int or TensorList of int) – Cropping window dimensions.
roi_start (int or list of int or TensorList of int) – ROI start coordinates.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
in_shape (int or list of int or TensorList of int, optional) –
Shape of the input data.

If provided, the cropping window start will be selected so that the cropping window is within the bounds of the input.

..note::
Providing in_shape is incompatible with feeding the input data directly as a positional input.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
roi_end (int or list of int or TensorList of int, optional) –
ROI end coordinates.

Note

Using roi_end is mutually exclusive with roi_shape.
roi_shape (int or list of int or TensorList of int, optional) –
ROI shape.

Note

Using roi_shape is mutually exclusive with roi_end.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.rotate(*inputs, **kwargs)¶

Rotates the images by the specified angle.

This operator supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList ('HWC', 'DHWC')) – Input to the operator.

Keyword Arguments

angle (float or TensorList of float) –
Angle, in degrees, by which the image is rotated.

For two-dimensional data, the rotation is counter-clockwise, assuming the top-left corner is at (0,0). For three-dimensional data, the angle is a positive rotation around the provided axis.
axis (float or list of float or TensorList of float, optional, default = []) –
Applies only to three-dimension and is the axis around which to rotate the image.

The vector does not need to be normalized, but it must have a non-zero length. Reversing the vector is equivalent to changing the sign of angle.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.NO_TYPE) –
Output data type.

If not set, the input type is used.
fill_value (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.FLOAT) –
Value used to fill areas that are outside the source image.

If a value is not specified, the source coordinates are clamped and the border pixel is repeated.
interp_type (nvidia.dali.types.DALIInterpType, optional, default = DALIInterpType.INTERP_LINEAR) – Type of interpolation used.
keep_size (bool, optional, default = False) –
If True, original canvas size is kept.

If set to False (default), and the size is not set, the canvas size is adjusted to accommodate the rotated image with the least padding possible.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
size (float or list of float or TensorList of float, optional, default = []) –
Output size, in pixels/points.

Non-integer sizes are rounded to nearest integer. The channel dimension should be excluded (for example, for RGB images, specify (480,640), not (480,640,3).
output_dtype (nvidia.dali.types.DALIDataType) –

Warning

The argument output_dtype is a deprecated alias for dtype. Use dtype instead.

nvidia.dali.fn.saturation(*inputs, **kwargs)¶

Changes the saturation level of the image.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList ('HWC', 'FHWC', 'DHWC')) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.UINT8) –
Output data type.

If not set, the input type is used.
image_type (nvidia.dali.types.DALIImageType, optional, default = DALIImageType.RGB) – The color space of the input and the output image.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
saturation (float or TensorList of float, optional, default = 1.0) –
The saturation change factor.

Values must be non-negative.

Example values:
- 0 - Completely desaturated image.
- 1 - No change to image’s saturation.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.sequence_reader(*inputs, **kwargs)¶: Warning

This operator is now deprecated. Use readers.sequence() instead.

In DALI 1.0 all readers were moved into a dedicated readers submodule and renamed to follow a common pattern. This is a placeholder operator with identical functionality to allow for backward compatibility.

Legacy alias for readers.sequence().

nvidia.dali.fn.sequence_rearrange(*inputs, **kwargs)¶

Rearranges frames in a sequence.

Assumes that the outermost dimension represents the frame index in the sequence. If the input has a non-empty layout description, it must start with F (frame).

This operator allows sequence inputs.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

new_order (int or list of int or TensorList of int) –
List that describes the new order for the elements in each sample.

Output sequence at position i will contain element new_order[i] from input sequence:
```
out[i, ...] = in[new_order[i], ...]
```
Elements can be repeated or dropped, but empty output sequences are not allowed. Only indices in [0, input_outermost_extent) are allowed to be used in new_order. Can be specified per sample as 1D tensors.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.shapes(*inputs, **kwargs)¶

Returns the shapes of inputs.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.INT64) – Data type to which the sizes are converted.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
type (nvidia.dali.types.DALIDataType) –

Warning

The argument type is a deprecated alias for dtype. Use dtype instead.

nvidia.dali.fn.slice(*inputs, **kwargs)¶

Extracts a subtensor, or slice.

Note

For generic indexing and slicing you can use Python indexing systax. See Indexing and Slicing for details.

The slice can be specified by proving the start and end coordinates, or start coordinates and shape of the slice. Both coordinates and shapes can be provided in absolute or relative terms.

The slice arguments can be specified by the following named arguments:

start: Slice start coordinates (absolute)
rel_start: Slice start coordinates (relative)
end: Slice end coordinates (absolute)
rel_end: Slice end coordinates (relative)
shape: Slice shape (absolute)
rel_shape: Slice shape (relative)

The slice can be configured by providing start and end coordinates or start and shape. Relative and absolute arguments can be mixed (for example, rel_start can be used with shape) as long as start and shape or end are uniquely defined.

Alternatively, two extra positional inputs can be provided, specifying anchor and shape. When using positional inputs, two extra boolean arguments normalized_anchor/normalized_shape can be used to specify the nature of the arguments provided. Using positional inputs for anchor and shape is incompatible with the named arguments specified above.

The slice arguments should provide as many dimensions as specified by the axis_names or axes arguments.

By default, the nvidia.dali.fn.slice() operator uses normalized coordinates and WH order for the slice arguments.

This operator supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

data (TensorList) – Batch that contains the input data.
anchor (1D TensorList of float or int, optional) –
(Optional) Input that contains normalized or absolute coordinates for the starting point of the slice (x0, x1, x2, …).

Integer coordinates are interpreted as absolute coordinates, while float coordinates can be interpreted as absolute or relative coordinates, depending on the value of normalized_anchor.
shape (1D TensorList of float or int, optional) –
(Optional) Input that contains normalized or absolute coordinates for the dimensions of the slice (s0, s1, s2, …).

Integer coordinates are interpreted as absolute coordinates, while float coordinates can be interpreted as absolute or relative coordinates, depending on the value of normalized_shape.

Keyword Arguments

axes (int or list of int or TensorList of int, optional, default = [1, 0]) –
Order of dimensions used for the anchor and shape slice inputs as dimension indices.

Negative values are interpreted as counting dimensions from the back. Valid range: [-ndim, ndim-1], where ndim is the number of dimensions in the input data.
axis_names (layout str, optional, default = ‘WH’) –
Order of the dimensions used for the anchor and shape slice inputs, as described in layout.

If a value is provided, axis_names will have a higher priority than axes.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.NO_TYPE) –
Output data type.

Supported types: FLOAT, FLOAT16, and UINT8.

If not set, the input type is used.
end (int or list of int or TensorList of int, optional) –
End coordinates of the slice.

Note: Providing named arguments start, end, shape, rel_start, rel_end, rel_shape is incompatible with providing positional inputs anchor and shape.
fill_values (float or list of float, optional, default = [0.0]) –
Determines padding values and is only relevant if out_of_bounds_policy is set to “pad”.

If a scalar value is provided, it will be used for all the channels. If multiple values are provided, the number of values and channels must be identical (extent of dimension C in the layout) in the output slice.
image_type (nvidia.dali.types.DALIImageType) –

Warning

The argument image_type is no longer used and will be removed in a future release.
normalized_anchor (bool, optional, default = True) –
Determines whether the anchor positional input should be interpreted as normalized (range [0.0, 1.0]) or as absolute coordinates.

Note

This argument is only relevant when anchor data type is float. For integer types, the coordinates are always absolute.
normalized_shape (bool, optional, default = True) –
Determines whether the shape positional input should be interpreted as normalized (range [0.0, 1.0]) or as absolute coordinates.

Note

This argument is only relevant when anchor data type is float. For integer types, the coordinates are always absolute.
out_of_bounds_policy (str, optional, default = ‘error’) –
Determines the policy when slicing the out of bounds area of the input.

Here is a list of the supported values:
- "error" (default): Attempting to slice outside of the bounds of the input will produce an error.
- "pad": The input will be padded as needed with zeros or any other value that is specified with the fill_values argument.
- "trim_to_shape": The slice window will be cut to the bounds of the input.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
rel_end (float or list of float or TensorList of float, optional) –
End relative coordinates of the slice (range [0.0 - 1.0].

Note: Providing named arguments start, end, shape, rel_start, rel_end, rel_shape is incompatible with providing positional inputs anchor and shape.
rel_shape (float or list of float or TensorList of float, optional) –
Relative shape of the slice (range [0.0 - 1.0]).

Providing named arguments start, end, shape, rel_start, rel_end, rel_shape is incompatible with providing positional inputs anchor and shape.
rel_start (float or list of float or TensorList of float, optional) –
Start relative coordinates of the slice (range [0.0 - 1.0]).

Note: Providing named arguments start, end, shape, rel_start, rel_end, rel_shape is incompatible with providing positional inputs anchor and shape.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shape (int or list of int or TensorList of int, optional) –
Shape of the slice.

Providing named arguments start, end, shape, rel_start, rel_end, rel_shape is incompatible with providing positional inputs anchor and shape.
start (int or list of int or TensorList of int, optional) –
Start coordinates of the slice.

Note: Providing named arguments start/end or start/shape is incompatible with providing positional inputs anchor and shape.
output_dtype (nvidia.dali.types.DALIDataType) –

Warning

The argument output_dtype is a deprecated alias for dtype. Use dtype instead.

nvidia.dali.fn.spectrogram(*inputs, **kwargs)¶

Produces a spectrogram from a 1D signal (for example, audio).

Input data is expected to be one channel (shape being (nsamples,), (nsamples, 1), or (1, nsamples)) of type float32.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
center_windows (bool, optional, default = True) –
Indicates whether extracted windows should be padded so that the window function is centered at multiples of window_step.

If set to False, the signal will not be padded, that is, only windows within the input range will be extracted.
layout (layout str, optional, default = ‘ft’) – Output layout: “ft” (frequency-major) or “tf” (time-major).
nfft (int, optional) –
Size of the FFT.

The number of bins that are created in the output is nfft // 2 + 1.

Note

The output only represents the positive part of the spectrum.
power (int, optional, default = 2) –
Exponent of the magnitude of the spectrum.

Supported values:
- 1 - amplitude,
- 2 - power (faster to compute).
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
reflect_padding (bool, optional, default = True) –
Indicates the padding policy when sampling outside the bounds of the signal.

If set to True, the signal is mirrored with respect to the boundary, otherwise the signal is padded with zeros.

Note

When center_windows is set to False, this option is ignored.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
window_fn (float or list of float, optional, default = []) –
Samples of the window function that will be multiplied to each extracted window when calculating the STFT.

If a value is provided, it should be a list of floating point numbers of size window_length. If a value is not provided, a Hann window will be used.
window_length (int, optional, default = 512) – Window size in number of samples.
window_step (int, optional, default = 256) – Step betweeen the STFT windows in number of samples.

nvidia.dali.fn.sphere(*inputs, **kwargs)¶

Performs a sphere augmentation.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList ('HWC')) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
fill_value (float, optional, default = 0.0) – Color value that is used for padding.
interp_type (nvidia.dali.types.DALIInterpType, optional, default = DALIInterpType.INTERP_NN) – Type of interpolation used.
mask (int or TensorList of int, optional, default = 1) –
Determines whether to apply this augmentation to the input image.

Here are the values:
- 0: Do not apply this transformation.
- 1: Apply this transformation.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.squeeze(*inputs, **kwargs)¶

Removes the dimensions given as axes or axis_names.

It’s an error to remove a dimension that would cause the total volume to change.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

data (TensorList) – Data to be squeezed

Keyword Arguments

axes (int or list of int or TensorList of int, optional, default = []) –
Indices of dimensions which should be removed.

All squeezed dimensions should have size 1, unless the total volume of the tensor is 0 before and after squeeze. All indices must be in the range of valid dimensions of the input
axis_names (layout str, optional, default = ‘’) –
Layout columns which should be removed.

All squeezed dimensions should have size 1, unless the total volume of the tensor is 0 before and after squeeze. All layout names should be present in data layout.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.ssd_random_crop(*inputs, **kwargs)¶

Warning

This operator is now deprecated. Use random_bbox_crop() instead.

Performs a random crop with bounding boxes where Intersection Over Union (IoU) meets a randomly selected threshold between 0-1.

When the IoU falls below the threshold, a new random crop is generated up to num_attempts. As an input, the operator accepts image, bounding boxes and labels. At the output cropped image, cropped and valid bounding boxes and valid labels are returned.

Supported backends

‘cpu’

Parameters

input0 (TensorList) – Input to the operator.
input1 (TensorList) – Input to the operator.
input2 (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
num_attempts (int, optional, default = 1) – Number of attempts.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.stack(*inputs, **kwargs)¶

Joins the input tensors along a new axis.

The shapes of respective tensors in the inputs must match.

Supported backends

‘cpu’
‘gpu’

Parameters

input0 (TensorList) – Input to the operator.
input[1..998] (TensorList, optional) – This function accepts up to 998 optional positional inputs

Keyword Arguments

axis (int, optional, default = 0) –
The axis in the output tensor along which the inputs are stacked.

The axis is inserted before a corresponding axis in the inputs. A value of 0 indicates that whole tensors are stacked. Specifying axis equal to the number of dimensions in the inputs causes the values from the inputs to be interleaved
axis_name (str, optional) –
Name of the new axis to be inserted.

A one-character string that will denote the new axis in the output layout. The output layout will be constructed by inserting that character into the input layout at the position indicated by axis. For example, specifying axis = 0 and axis_name = "C" with input layout “HW” will yield the output layout “CHW”
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.tfrecord_reader(*inputs, **kwargs)¶: Warning

This operator is now deprecated. Use readers.tfrecord() instead.

In DALI 1.0 all readers were moved into a dedicated readers submodule and renamed to follow a common pattern. This is a placeholder operator with identical functionality to allow for backward compatibility.

Legacy alias for readers.tfrecord().

nvidia.dali.fn.to_decibels(*inputs, **kwargs)¶

Converts a magnitude (real, positive) to the decibel scale by using the following formula:

min_ratio = pow(10, cutoff_db / multiplier)
out[i] = multiplier * log10( max(min_ratio, input[i] / reference) )

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
cutoff_db (float, optional, default = -200.0) –
Minimum or cut-off ratio in dB.

Any value below this value will saturate. For example, a value of cutoff_db=-80 corresponds to a minimum ratio of 1e-8.
multiplier (float, optional, default = 10.0) – Factor by which the logarithm is multiplied. The value is typically 10.0 or 20.0, which depends on whether the magnitude is squared.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
reference (float, optional, default = 0.0) –
Reference magnitude.

If a value is not provided, the maximum value for the input will be used as reference.

Note

The maximum of the input will be calculated on a per-sample basis.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.transpose(*inputs, **kwargs)¶

Transposes the tensors by reordering the dimensions based on the perm parameter.

Destination dimension i is obtained from source dimension perm[i].

For example, for a source image with HWC layout, shape = (100, 200, 3), and perm = [2, 0, 1], it will produce a destination image with CHW layout and shape = (3, 100, 200), holding the equality:

\[dst(x_2, x_0, x_1) = src(x_0, x_1, x_2)\]

which is equivalent to:

\[dst(x_{perm[0]}, x_{perm[1]}, x_{perm[2]}) = src(x_0, x_1, x_2)\]

for all valid coordinates.

This operator allows sequence inputs and supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

perm (int or list of int) – Permutation of the dimensions of the input, for example, [2, 0, 1].
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
output_layout (layout str, optional, default = ‘’) –
Explicitly sets the output data layout.

If this argument is specified, transpose_layout is ignored.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
transpose_layout (bool, optional, default = True) –
When set to True, the axis names in the output data layout are permuted according to perm, Otherwise, the input layout is copied to the output.

If output_layout is set, this argument is ignored.

nvidia.dali.fn.uniform(*inputs, **kwargs)¶

Warning

This operator is now deprecated. Use random.uniform() instead.

Generates random numbers following a uniform distribution.

It can be configured to produce a continuous uniform distribution in the range [min, max), or a discrete uniform distribution where any of the specified values [v0, v1, …, vn] occur with equal probability.

The shape of the generated data can be either specified explicitly with a shape argument, or chosen to match the shape of the input, if provided. If none are present, a scalar is generated.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList, optional) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional) –
Output data type.

Note

The generated numbers are converted to the output data type, rounding and clamping if necessary.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
range (float or list of float or TensorList of float, optional, default = [-1.0, 1.0]) –
Range [min, max) of a continuous uniform distribution.

This argument is mutually exclusive with values.

Warning

When specifying an integer type as dtype, the generated numbers can go outside the specified range, due to rounding.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shape (int or list of int or TensorList of int, optional) – Shape of the output data.
values (float or list of float or TensorList of float, optional) –
The discrete values [v0, v1, …, vn] produced by a discrete uniform distribution.

This argument is mutually exclusive with range.

nvidia.dali.fn.video_reader(*inputs, **kwargs)¶: Warning

This operator is now deprecated. Use readers.video() instead.

In DALI 1.0 all readers were moved into a dedicated readers submodule and renamed to follow a common pattern. This is a placeholder operator with identical functionality to allow for backward compatibility.

Legacy alias for readers.video().

nvidia.dali.fn.video_reader_resize(*inputs, **kwargs)¶: Warning

This operator is now deprecated. Use readers.video_resize() instead.

In DALI 1.0 all readers were moved into a dedicated readers submodule and renamed to follow a common pattern. This is a placeholder operator with identical functionality to allow for backward compatibility.

Legacy alias for readers.video_resize().

nvidia.dali.fn.warp_affine(*inputs, **kwargs)¶

Applies an affine transformation to the images.

This operator supports volumetric data.

Supported backends

‘cpu’
‘gpu’

Parameters

input0 (TensorList ('HWC', 'DHWC')) – Input to the operator.
input1 (TensorList, optional) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.NO_TYPE) –
Output data type.

If not set, the input type is used.
fill_value (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.FLOAT) –
Value used to fill areas that are outside the source image.

If a value is not specified, the source coordinates are clamped and the border pixel is repeated.
interp_type (nvidia.dali.types.DALIInterpType, optional, default = DALIInterpType.INTERP_LINEAR) – Type of interpolation used.
inverse_map (bool, optional, default = True) – Set to False if the given transform is a destination to source mapping, True otherwise.
matrix (float or list of float or TensorList of float, optional, default = []) –
Transform matrix.

When the inverse_map option is set to true (default), the matrix represents a destination to source mapping. With a list of values (M11, M12, M13, M21, M22, M23), this operation produces a new image by using the following formula:
```
dst(x,y) = src(M11 * x + M12 * y + M13, M21 * x + M22 * y + M23)
```
Where [0, 0] coordinate means the corner of the first pixel.

If the inverse_map option is set to false, the matrix represents a source to destination transform and it is inverted before applying the formula above.

It is equivalent to OpenCV’s warpAffine operation with the inverse_map argument being analog to the WARP_INVERSE_MAP flag.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
size (float or list of float or TensorList of float, optional, default = []) –
Output size, in pixels/points.

Non-integer sizes are rounded to nearest integer. The channel dimension should be excluded (for example, for RGB images, specify (480,640), not (480,640,3).
output_dtype (nvidia.dali.types.DALIDataType) –

Warning

The argument output_dtype is a deprecated alias for dtype. Use dtype instead.

nvidia.dali.fn.water(*inputs, **kwargs)¶

Performs a water augmentation, which makes the image appear to be underwater.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList ('HWC')) – Input to the operator.

Keyword Arguments

ampl_x (float, optional, default = 10.0) – Amplitude of the wave in the x direction.
ampl_y (float, optional, default = 10.0) – Amplitude of the wave in the y direction.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
fill_value (float, optional, default = 0.0) – Color value that is used for padding.
freq_x (float, optional, default = 0.049087) – Frequency of the wave in the x direction.
freq_y (float, optional, default = 0.049087) – Frequence of the wave in the y direction.
interp_type (nvidia.dali.types.DALIInterpType, optional, default = DALIInterpType.INTERP_NN) – Type of interpolation used.
mask (int or TensorList of int, optional, default = 1) –
Determines whether to apply this augmentation to the input image.

Here are the values:
- 0: Do not apply this transformation.
- 1: Apply this transformation.
phase_x (float, optional, default = 0.0) – Phase of the wave in the x direction.
phase_y (float, optional, default = 0.0) – Phase of the wave in the y direction.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.decoders¶

nvidia.dali.fn.decoders.audio(*inputs, **kwargs)¶

Decodes waveforms from encoded audio data.

It supports the following audio formats: wav, flac and ogg. This operator produces the following outputs:

output[0]: A batch of decoded data
output[1]: A batch of sampling rates [Hz].

Supported backends

‘cpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
downmix (bool, optional, default = False) –
If set to True, downmix all input channels to mono.

If downmixing is turned on, the decoder output is 1D. If downmixing is turned off, it produces 2D output with interleaved channels.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.FLOAT) –
Output data type.

Supported types: INT16, INT32, FLOAT.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
quality (float, optional, default = 50.0) –
Resampling quality, where 0 is the lowest, and 100 is the highest.

0 gives 3 lobes of the sinc filter, 50 gives 16 lobes, and 100 gives 64 lobes.
sample_rate (float or TensorList of float, optional, default = 0.0) – If specified, the target sample rate, in Hz, to which the audio is resampled.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.decoders.image(*inputs, **kwargs)¶

Decodes images.

For jpeg images, depending on the backend selected (“mixed” and “cpu”), the implementation uses the nvJPEG library or libjpeg-turbo, respectively. Other image formats are decoded with OpenCV or other specific libraries, such as libtiff.

If used with a mixed backend, and the hardware is available, the operator will use a dedicated hardware decoder.

Warning

Due to performance reasons, hardware decoder is disabled for driver < 455.x

The output of the decoder is in HWC layout.

Supported formats: JPG, BMP, PNG, TIFF, PNM, PPM, PGM, PBM, JPEG 2000, WebP. Please note that GPU acceleration for JPEG 2000 decoding is only available for CUDA 11.

Note

WebP decoding currently only supports the simple file format (lossy and lossless compression). For details on the different WebP file formats, see https://developers.google.com/speed/webp/docs/riff_container

Note

EXIF orientation metadata is disregarded.

Supported backends

‘cpu’
‘mixed’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

affine (bool, optional, default = True) –
Applies only to the mixed backend type.

If set to True, each thread in the internal thread pool will be tied to a specific CPU core. Otherwise, the threads can be reassigned to any CPU core by the operating system.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
cache_batch_copy (bool, optional, default = True) –
Applies only to the mixed backend type.

If set to True, multiple images from the cache are copied with a batched copy kernel call. Otherwise, unless the order in the batch is the same as in the cache, each image is copied with cudaMemcpy.
cache_debug (bool, optional, default = False) –
Applies only to the mixed backend type.

Prints the debug information about the decoder cache.
cache_size (int, optional, default = 0) –
Applies only to the mixed backend type.

Total size of the decoder cache in megabytes. When provided, the decoded images that are larger than cache_threshold will be cached in GPU memory.
cache_threshold (int, optional, default = 0) –
Applies only to the mixed backend type.

The size threshold, in bytes, for decoded images to be cached. When an image is cached, it no longer needs to be decoded when it is encountered at the operator input saving processing time.
cache_type (str, optional, default = ‘’) –
Applies only to the mixed backend type.

Here is a list of the available cache types:
- threshold: caches every image with a size that is larger than cache_threshold until
  
  the cache is full.
  
  The warm-up time for threshold policy is 1 epoch.
- largest: stores the largest images that can fit in the cache.
  
  The warm-up time for largest policy is 2 epochs
  
  Note
  
  To take advantage of caching, it is recommended to configure readers with stick_to_shard=True to limit the amount of unique images seen by each decoder instance in a multi node environment.
device_memory_padding (int, optional, default = 16777216) –
Applies only to the mixed backend type.

The padding for nvJPEG’s device memory allocations, in bytes. This parameter helps to avoid reallocation in nvJPEG when a larger image is encountered, and the internal buffer needs to be reallocated to decode the image.

If a value greater than 0 is provided, the operator preallocates one device buffer of the requested size per thread. If the value is correctly selected, no additional allocations will occur during the pipeline execution. One way to find the ideal value is to do a complete run over the dataset with the memory_stats argument set to True and then copy the largest allocation value that was printed in the statistics.
device_memory_padding_jpeg2k (int, optional, default = 0) –
Applies only to the mixed backend type.

The padding for nvJPEG2k’s device memory allocations, in bytes. This parameter helps to avoid reallocation in nvJPEG2k when a larger image is encountered, and the internal buffer needs to be reallocated to decode the image.

If a value greater than 0 is provided, the operator preallocates the necessary number of buffers according to the hint provided. If the value is correctly selected, no additional allocations will occur during the pipeline execution. One way to find the ideal value is to do a complete run over the dataset with the memory_stats argument set to True and then copy the largest allocation value that was printed in the statistics.
host_memory_padding (int, optional, default = 8388608) –
Applies only to the mixed backend type.

The padding for nvJPEG’s host memory allocations, in bytes. This parameter helps to prevent the reallocation in nvJPEG when a larger image is encountered, and the internal buffer needs to be reallocated to decode the image.

If a value greater than 0 is provided, the operator preallocates two (because of double-buffering) host-pinned buffers of the requested size per thread. If selected correctly, no additional allocations will occur during the pipeline execution. One way to find the ideal value is to do a complete run over the dataset with the memory_stats argument set to True, and then copy the largest allocation value that is printed in the statistics.
host_memory_padding_jpeg2k (int, optional, default = 0) –
Applies only to the mixed backend type.

The padding for nvJPEG2k’s host memory allocations, in bytes. This parameter helps to prevent the reallocation in nvJPEG2k when a larger image is encountered, and the internal buffer needs to be reallocated to decode the image.

If a value greater than 0 is provided, the operator preallocates the necessary number of buffers according to the hint provided. If the value is correctly selected, no additional allocations will occur during the pipeline execution. One way to find the ideal value is to do a complete run over the dataset with the memory_stats argument set to True, and then copy the largest allocation value that is printed in the statistics.
hw_decoder_load (float, optional, default = 0.65) –
The percentage of the image data to be processed by the HW JPEG decoder.

Applies only to the mixed backend type in NVIDIA Ampere GPU architecture.

Determines the percentage of the workload that will be offloaded to the hardware decoder, if available. The optimal workload depends on the number of threads that are provided to the DALI pipeline and should be found empirically. More details can be found at https://developer.nvidia.com/blog/loading-data-fast-with-dali-and-new-jpeg-decoder-in-a100
hybrid_huffman_threshold (int, optional, default = 1000000) –
Applies only to the mixed backend type.

Images with a total number of pixels (height * width) that is higher than this threshold will use the nvJPEG hybrid Huffman decoder. Images that have fewer pixels will use the nvJPEG host-side Huffman decoder.

Note

Hybrid Huffman decoder still largely uses the CPU.
memory_stats (bool, optional, default = False) –
Applies only to the mixed backend type.

Prints debug information about nvJPEG allocations. The information about the largest allocation might be useful to determine suitable values for device_memory_padding and host_memory_padding for a dataset.

Note

The statistics are global for the entire process, not per operator instance, and include the allocations made during construction if the padding hints are non-zero.
output_type (nvidia.dali.types.DALIImageType, optional, default = DALIImageType.RGB) –
The color space of the output image.

Note: When decoding to YCbCr, the image will be decoded to RGB and then converted to YCbCr, following the YCbCr definition from ITU-R BT.601.
preallocate_height_hint (int, optional, default = 0) –
Image width hint.

Applies only to the mixed backend type in NVIDIA Ampere GPU architecture.

The hint is used to preallocate memory for the HW JPEG decoder.
preallocate_width_hint (int, optional, default = 0) –
Image width hint.

Applies only to the mixed backend type in NVIDIA Ampere GPU architecture.

The hint is used to preallocate memory for the HW JPEG decoder.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
split_stages (bool) –

Warning

The argument split_stages is no longer used and will be removed in a future release.
use_chunk_allocator (bool) –

Warning

The argument use_chunk_allocator is no longer used and will be removed in a future release.
use_fast_idct (bool, optional, default = False) –
Enables fast IDCT in the libjpeg-turbo based CPU decoder, used when device is set to “cpu” or when the it is set to “mixed” but the particular image can not be handled by the GPU implementation.

According to the libjpeg-turbo documentation, decompression performance is improved by up to 14% with little reduction in quality.

nvidia.dali.fn.decoders.image_crop(*inputs, **kwargs)¶

Decodes images and extracts regions-of-interest (ROI) that are specified by fixed window dimensions and variable anchors.

When possible, the argument uses the ROI decoding APIs (for example, libjpeg-turbo and nvJPEG) to reduce the decoding time and memory usage. When the ROI decoding is not supported for a given image format, it will decode the entire image and crop the selected ROI.

The output of the decoder is in HWC layout.

Supported formats: JPG, BMP, PNG, TIFF, PNM, PPM, PGM, PBM, JPEG 2000, WebP.

Note

JPEG 2000 region-of-interest (ROI) decoding is not accelerated on the GPU, and will use a CPU implementation regardless of the selected backend. For a GPU accelerated implementation, consider using separate decoders.image and crop operators.

Note

EXIF orientation metadata is disregarded.

Supported backends

‘cpu’
‘mixed’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

affine (bool, optional, default = True) –
Applies only to the mixed backend type.

If set to True, each thread in the internal thread pool will be tied to a specific CPU core. Otherwise, the threads can be reassigned to any CPU core by the operating system.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
crop (float or list of float, optional, default = [0.0, 0.0]) –
Shape of the cropped image, specified as a list of values (for example, (crop_H, crop_W) for the 2D crop and (crop_D, crop_H, crop_W) for the volumetric crop).

Providing crop argument is incompatible with providing separate arguments such as crop_d, crop_h, and crop_w.
crop_d (float or TensorList of float, optional, default = 0.0) –
Applies only to volumetric inputs; cropping window depth (in voxels).

crop_w, crop_h, and crop_d must be specified together. Providing values for crop_w, crop_h, and crop_d is incompatible with providing the fixed crop window dimensions (argument crop).
crop_h (float or TensorList of float, optional, default = 0.0) –
Cropping the window height (in pixels).

Providing values for crop_w and crop_h is incompatible with providing fixed crop window dimensions (argument crop).
crop_pos_x (float or TensorList of float, optional, default = 0.5) –
Normalized (0.0 - 1.0) horizontal position of the cropping window (upper left corner).

The actual position is calculated as crop_x = crop_x_norm * (W - crop_W), where crop_x_norm is the normalized position, W is the width of the image, and crop_W is the width of the cropping window.
crop_pos_y (float or TensorList of float, optional, default = 0.5) –
Normalized (0.0 - 1.0) vertical position of the start of the cropping window (typically, the upper left corner).

The actual position is calculated as crop_y = crop_y_norm * (H - crop_H), where crop_y_norm is the normalized position, H is the height of the image, and crop_H is the height of the cropping window.
crop_pos_z (float or TensorList of float, optional, default = 0.5) –
Applies only to volumetric inputs.

Normalized (0.0 - 1.0) normal position of the cropping window (front plane). The actual position is calculated as crop_z = crop_z_norm * (D - crop_D), where crop_z_norm is the normalized position, D is the depth of the image and crop_D is the depth of the cropping window.
crop_w (float or TensorList of float, optional, default = 0.0) –
Cropping window width (in pixels).

Providing values for crop_w and crop_h is incompatible with providing fixed crop window dimensions (argument crop).
device_memory_padding (int, optional, default = 16777216) –
Applies only to the mixed backend type.

The padding for nvJPEG’s device memory allocations, in bytes. This parameter helps to avoid reallocation in nvJPEG when a larger image is encountered, and the internal buffer needs to be reallocated to decode the image.

If a value greater than 0 is provided, the operator preallocates one device buffer of the requested size per thread. If the value is correctly selected, no additional allocations will occur during the pipeline execution. One way to find the ideal value is to do a complete run over the dataset with the memory_stats argument set to True and then copy the largest allocation value that was printed in the statistics.
device_memory_padding_jpeg2k (int, optional, default = 0) –
Applies only to the mixed backend type.

The padding for nvJPEG2k’s device memory allocations, in bytes. This parameter helps to avoid reallocation in nvJPEG2k when a larger image is encountered, and the internal buffer needs to be reallocated to decode the image.

If a value greater than 0 is provided, the operator preallocates the necessary number of buffers according to the hint provided. If the value is correctly selected, no additional allocations will occur during the pipeline execution. One way to find the ideal value is to do a complete run over the dataset with the memory_stats argument set to True and then copy the largest allocation value that was printed in the statistics.
host_memory_padding (int, optional, default = 8388608) –
Applies only to the mixed backend type.

The padding for nvJPEG’s host memory allocations, in bytes. This parameter helps to prevent the reallocation in nvJPEG when a larger image is encountered, and the internal buffer needs to be reallocated to decode the image.

If a value greater than 0 is provided, the operator preallocates two (because of double-buffering) host-pinned buffers of the requested size per thread. If selected correctly, no additional allocations will occur during the pipeline execution. One way to find the ideal value is to do a complete run over the dataset with the memory_stats argument set to True, and then copy the largest allocation value that is printed in the statistics.
host_memory_padding_jpeg2k (int, optional, default = 0) –
Applies only to the mixed backend type.

The padding for nvJPEG2k’s host memory allocations, in bytes. This parameter helps to prevent the reallocation in nvJPEG2k when a larger image is encountered, and the internal buffer needs to be reallocated to decode the image.

If a value greater than 0 is provided, the operator preallocates the necessary number of buffers according to the hint provided. If the value is correctly selected, no additional allocations will occur during the pipeline execution. One way to find the ideal value is to do a complete run over the dataset with the memory_stats argument set to True, and then copy the largest allocation value that is printed in the statistics.
hw_decoder_load (float, optional, default = 0.65) –
The percentage of the image data to be processed by the HW JPEG decoder.

Applies only to the mixed backend type in NVIDIA Ampere GPU architecture.

Determines the percentage of the workload that will be offloaded to the hardware decoder, if available. The optimal workload depends on the number of threads that are provided to the DALI pipeline and should be found empirically. More details can be found at https://developer.nvidia.com/blog/loading-data-fast-with-dali-and-new-jpeg-decoder-in-a100
hybrid_huffman_threshold (int, optional, default = 1000000) –
Applies only to the mixed backend type.

Images with a total number of pixels (height * width) that is higher than this threshold will use the nvJPEG hybrid Huffman decoder. Images that have fewer pixels will use the nvJPEG host-side Huffman decoder.

Note

Hybrid Huffman decoder still largely uses the CPU.
memory_stats (bool, optional, default = False) –
Applies only to the mixed backend type.

Prints debug information about nvJPEG allocations. The information about the largest allocation might be useful to determine suitable values for device_memory_padding and host_memory_padding for a dataset.

Note

The statistics are global for the entire process, not per operator instance, and include the allocations made during construction if the padding hints are non-zero.
output_type (nvidia.dali.types.DALIImageType, optional, default = DALIImageType.RGB) –
The color space of the output image.

Note: When decoding to YCbCr, the image will be decoded to RGB and then converted to YCbCr, following the YCbCr definition from ITU-R BT.601.
preallocate_height_hint (int, optional, default = 0) –
Image width hint.

Applies only to the mixed backend type in NVIDIA Ampere GPU architecture.

The hint is used to preallocate memory for the HW JPEG decoder.
preallocate_width_hint (int, optional, default = 0) –
Image width hint.

Applies only to the mixed backend type in NVIDIA Ampere GPU architecture.

The hint is used to preallocate memory for the HW JPEG decoder.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
split_stages (bool) –

Warning

The argument split_stages is no longer used and will be removed in a future release.
use_chunk_allocator (bool) –

Warning

The argument use_chunk_allocator is no longer used and will be removed in a future release.
use_fast_idct (bool, optional, default = False) –
Enables fast IDCT in the libjpeg-turbo based CPU decoder, used when device is set to “cpu” or when the it is set to “mixed” but the particular image can not be handled by the GPU implementation.

According to the libjpeg-turbo documentation, decompression performance is improved by up to 14% with little reduction in quality.

nvidia.dali.fn.decoders.image_random_crop(*inputs, **kwargs)¶

Decodes images and randomly crops them.

The cropping window’s area (relative to the entire image) and aspect ratio can be restricted to a range of values specified by area and aspect_ratio arguments, respectively.

When possible, the operator uses the ROI decoding APIs (for example, libjpeg-turbo and nvJPEG) to reduce the decoding time and memory usage. When the ROI decoding is not supported for a given image format, it will decode the entire image and crop the selected ROI.

The output of the decoder is in HWC layout.

Supported formats: JPG, BMP, PNG, TIFF, PNM, PPM, PGM, PBM, JPEG 2000, WebP.

Note

JPEG 2000 region-of-interest (ROI) decoding is not accelerated on the GPU, and will use a CPU implementation regardless of the selected backend. For a GPU accelerated implementation, consider using separate decoders.image and random_crop operators.

Note

EXIF orientation metadata is disregarded.

Supported backends

‘cpu’
‘mixed’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

affine (bool, optional, default = True) –
Applies only to the mixed backend type.

If set to True, each thread in the internal thread pool will be tied to a specific CPU core. Otherwise, the threads can be reassigned to any CPU core by the operating system.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
device_memory_padding (int, optional, default = 16777216) –
Applies only to the mixed backend type.

The padding for nvJPEG’s device memory allocations, in bytes. This parameter helps to avoid reallocation in nvJPEG when a larger image is encountered, and the internal buffer needs to be reallocated to decode the image.

If a value greater than 0 is provided, the operator preallocates one device buffer of the requested size per thread. If the value is correctly selected, no additional allocations will occur during the pipeline execution. One way to find the ideal value is to do a complete run over the dataset with the memory_stats argument set to True and then copy the largest allocation value that was printed in the statistics.
device_memory_padding_jpeg2k (int, optional, default = 0) –
Applies only to the mixed backend type.

The padding for nvJPEG2k’s device memory allocations, in bytes. This parameter helps to avoid reallocation in nvJPEG2k when a larger image is encountered, and the internal buffer needs to be reallocated to decode the image.

If a value greater than 0 is provided, the operator preallocates the necessary number of buffers according to the hint provided. If the value is correctly selected, no additional allocations will occur during the pipeline execution. One way to find the ideal value is to do a complete run over the dataset with the memory_stats argument set to True and then copy the largest allocation value that was printed in the statistics.
host_memory_padding (int, optional, default = 8388608) –
Applies only to the mixed backend type.

The padding for nvJPEG’s host memory allocations, in bytes. This parameter helps to prevent the reallocation in nvJPEG when a larger image is encountered, and the internal buffer needs to be reallocated to decode the image.

If a value greater than 0 is provided, the operator preallocates two (because of double-buffering) host-pinned buffers of the requested size per thread. If selected correctly, no additional allocations will occur during the pipeline execution. One way to find the ideal value is to do a complete run over the dataset with the memory_stats argument set to True, and then copy the largest allocation value that is printed in the statistics.
host_memory_padding_jpeg2k (int, optional, default = 0) –
Applies only to the mixed backend type.

The padding for nvJPEG2k’s host memory allocations, in bytes. This parameter helps to prevent the reallocation in nvJPEG2k when a larger image is encountered, and the internal buffer needs to be reallocated to decode the image.

If a value greater than 0 is provided, the operator preallocates the necessary number of buffers according to the hint provided. If the value is correctly selected, no additional allocations will occur during the pipeline execution. One way to find the ideal value is to do a complete run over the dataset with the memory_stats argument set to True, and then copy the largest allocation value that is printed in the statistics.
hw_decoder_load (float, optional, default = 0.65) –
The percentage of the image data to be processed by the HW JPEG decoder.

Applies only to the mixed backend type in NVIDIA Ampere GPU architecture.

Determines the percentage of the workload that will be offloaded to the hardware decoder, if available. The optimal workload depends on the number of threads that are provided to the DALI pipeline and should be found empirically. More details can be found at https://developer.nvidia.com/blog/loading-data-fast-with-dali-and-new-jpeg-decoder-in-a100
hybrid_huffman_threshold (int, optional, default = 1000000) –
Applies only to the mixed backend type.

Images with a total number of pixels (height * width) that is higher than this threshold will use the nvJPEG hybrid Huffman decoder. Images that have fewer pixels will use the nvJPEG host-side Huffman decoder.

Note

Hybrid Huffman decoder still largely uses the CPU.
memory_stats (bool, optional, default = False) –
Applies only to the mixed backend type.

Prints debug information about nvJPEG allocations. The information about the largest allocation might be useful to determine suitable values for device_memory_padding and host_memory_padding for a dataset.

Note

The statistics are global for the entire process, not per operator instance, and include the allocations made during construction if the padding hints are non-zero.
num_attempts (int, optional, default = 10) – Maximum number of attempts used to choose random area and aspect ratio.
output_type (nvidia.dali.types.DALIImageType, optional, default = DALIImageType.RGB) –
The color space of the output image.

Note: When decoding to YCbCr, the image will be decoded to RGB and then converted to YCbCr, following the YCbCr definition from ITU-R BT.601.
preallocate_height_hint (int, optional, default = 0) –
Image width hint.

Applies only to the mixed backend type in NVIDIA Ampere GPU architecture.

The hint is used to preallocate memory for the HW JPEG decoder.
preallocate_width_hint (int, optional, default = 0) –
Image width hint.

Applies only to the mixed backend type in NVIDIA Ampere GPU architecture.

The hint is used to preallocate memory for the HW JPEG decoder.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
random_area (float or list of float, optional, default = [0.08, 1.0]) –
Range from which to choose random area fraction A.

The cropped image’s area will be equal to A * original image’s area.
random_aspect_ratio (float or list of float, optional, default = [0.75, 1.333333]) – Range from which to choose random aspect ratio (width/height).
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
split_stages (bool) –

Warning

The argument split_stages is no longer used and will be removed in a future release.
use_chunk_allocator (bool) –

Warning

The argument use_chunk_allocator is no longer used and will be removed in a future release.
use_fast_idct (bool, optional, default = False) –
Enables fast IDCT in the libjpeg-turbo based CPU decoder, used when device is set to “cpu” or when the it is set to “mixed” but the particular image can not be handled by the GPU implementation.

According to the libjpeg-turbo documentation, decompression performance is improved by up to 14% with little reduction in quality.

nvidia.dali.fn.decoders.image_slice(*inputs, **kwargs)¶

Decodes images and extracts regions of interest.

The slice can be specified by proving the start and end coordinates, or start coordinates and shape of the slice. Both coordinates and shapes can be provided in absolute or relative terms.

The slice arguments can be specified by the following named arguments:

start: Slice start coordinates (absolute)
rel_start: Slice start coordinates (relative)
end: Slice end coordinates (absolute)
rel_end: Slice end coordinates (relative)
shape: Slice shape (absolute)
rel_shape: Slice shape (relative)

The slice can be configured by providing start and end coordinates or start and shape. Relative and absolute arguments can be mixed (for example, rel_start can be used with shape) as long as start and shape or end are uniquely defined.

Alternatively, two extra positional inputs can be provided, specifying anchor and shape. When using positional inputs, two extra boolean arguments normalized_anchor/normalized_shape can be used to specify the nature of the arguments provided. Using positional inputs for anchor and shape is incompatible with the named arguments specified above.

The slice arguments should provide as many dimensions as specified by the axis_names or axes arguments.

By default, the nvidia.dali.fn.decoders.image_slice() operator uses normalized coordinates and “WH” order for the slice arguments.

When possible, the argument uses the ROI decoding APIs (for example, libjpeg-turbo and nvJPEG) to optimize the decoding time and memory usage. When the ROI decoding is not supported for a given image format, it will decode the entire image and crop the selected ROI.

The output of the decoder is in the HWC layout.

Supported formats: JPG, BMP, PNG, TIFF, PNM, PPM, PGM, PBM, JPEG 2000, WebP.

Note

JPEG 2000 region-of-interest (ROI) decoding is not accelerated on the GPU, and will use a CPU implementation regardless of the selected backend. For a GPU accelerated implementation, consider using separate decoders.image and slice operators.

Note

EXIF orientation metadata is disregarded.

Supported backends

‘cpu’
‘mixed’

Parameters

data (TensorList) – Batch that contains the input data.
anchor (1D TensorList of float or int, optional) –
Input that contains normalized or absolute coordinates for the starting point of the slice (x0, x1, x2, …).

Integer coordinates are interpreted as absolute coordinates, while float coordinates can be interpreted as absolute or relative coordinates, depending on the value of normalized_anchor.
shape (1D TensorList of float or int, optional) –
Input that contains normalized or absolute coordinates for the dimensions of the slice (s0, s1, s2, …).

Integer coordinates are interpreted as absolute coordinates, while float coordinates can be interpreted as absolute or relative coordinates, depending on the value of normalized_shape.

Keyword Arguments

affine (bool, optional, default = True) –
Applies only to the mixed backend type.

If set to True, each thread in the internal thread pool will be tied to a specific CPU core. Otherwise, the threads can be reassigned to any CPU core by the operating system.
axes (int or list of int or TensorList of int, optional, default = [1, 0]) –
Order of dimensions used for the anchor and shape slice inputs as dimension indices.

Negative values are interpreted as counting dimensions from the back. Valid range: [-ndim, ndim-1], where ndim is the number of dimensions in the input data.
axis_names (layout str, optional, default = ‘WH’) –
Order of the dimensions used for the anchor and shape slice inputs, as described in layout.

If a value is provided, axis_names will have a higher priority than axes.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
device_memory_padding (int, optional, default = 16777216) –
Applies only to the mixed backend type.

The padding for nvJPEG’s device memory allocations, in bytes. This parameter helps to avoid reallocation in nvJPEG when a larger image is encountered, and the internal buffer needs to be reallocated to decode the image.

If a value greater than 0 is provided, the operator preallocates one device buffer of the requested size per thread. If the value is correctly selected, no additional allocations will occur during the pipeline execution. One way to find the ideal value is to do a complete run over the dataset with the memory_stats argument set to True and then copy the largest allocation value that was printed in the statistics.
device_memory_padding_jpeg2k (int, optional, default = 0) –
Applies only to the mixed backend type.

The padding for nvJPEG2k’s device memory allocations, in bytes. This parameter helps to avoid reallocation in nvJPEG2k when a larger image is encountered, and the internal buffer needs to be reallocated to decode the image.

If a value greater than 0 is provided, the operator preallocates the necessary number of buffers according to the hint provided. If the value is correctly selected, no additional allocations will occur during the pipeline execution. One way to find the ideal value is to do a complete run over the dataset with the memory_stats argument set to True and then copy the largest allocation value that was printed in the statistics.
end (int or list of int or TensorList of int, optional) –
End coordinates of the slice.

Note: Providing named arguments start, end, shape, rel_start, rel_end, rel_shape is incompatible with providing positional inputs anchor and shape.
host_memory_padding (int, optional, default = 8388608) –
Applies only to the mixed backend type.

The padding for nvJPEG’s host memory allocations, in bytes. This parameter helps to prevent the reallocation in nvJPEG when a larger image is encountered, and the internal buffer needs to be reallocated to decode the image.

If a value greater than 0 is provided, the operator preallocates two (because of double-buffering) host-pinned buffers of the requested size per thread. If selected correctly, no additional allocations will occur during the pipeline execution. One way to find the ideal value is to do a complete run over the dataset with the memory_stats argument set to True, and then copy the largest allocation value that is printed in the statistics.
host_memory_padding_jpeg2k (int, optional, default = 0) –
Applies only to the mixed backend type.

The padding for nvJPEG2k’s host memory allocations, in bytes. This parameter helps to prevent the reallocation in nvJPEG2k when a larger image is encountered, and the internal buffer needs to be reallocated to decode the image.

If a value greater than 0 is provided, the operator preallocates the necessary number of buffers according to the hint provided. If the value is correctly selected, no additional allocations will occur during the pipeline execution. One way to find the ideal value is to do a complete run over the dataset with the memory_stats argument set to True, and then copy the largest allocation value that is printed in the statistics.
hw_decoder_load (float, optional, default = 0.65) –
The percentage of the image data to be processed by the HW JPEG decoder.

Applies only to the mixed backend type in NVIDIA Ampere GPU architecture.

Determines the percentage of the workload that will be offloaded to the hardware decoder, if available. The optimal workload depends on the number of threads that are provided to the DALI pipeline and should be found empirically. More details can be found at https://developer.nvidia.com/blog/loading-data-fast-with-dali-and-new-jpeg-decoder-in-a100
hybrid_huffman_threshold (int, optional, default = 1000000) –
Applies only to the mixed backend type.

Images with a total number of pixels (height * width) that is higher than this threshold will use the nvJPEG hybrid Huffman decoder. Images that have fewer pixels will use the nvJPEG host-side Huffman decoder.

Note

Hybrid Huffman decoder still largely uses the CPU.
memory_stats (bool, optional, default = False) –
Applies only to the mixed backend type.

Prints debug information about nvJPEG allocations. The information about the largest allocation might be useful to determine suitable values for device_memory_padding and host_memory_padding for a dataset.

Note

The statistics are global for the entire process, not per operator instance, and include the allocations made during construction if the padding hints are non-zero.
normalized_anchor (bool, optional, default = True) –
Determines whether the anchor positional input should be interpreted as normalized (range [0.0, 1.0]) or as absolute coordinates.

Note

This argument is only relevant when anchor data type is float. For integer types, the coordinates are always absolute.
normalized_shape (bool, optional, default = True) –
Determines whether the shape positional input should be interpreted as normalized (range [0.0, 1.0]) or as absolute coordinates.

Note

This argument is only relevant when anchor data type is float. For integer types, the coordinates are always absolute.
output_type (nvidia.dali.types.DALIImageType, optional, default = DALIImageType.RGB) –
The color space of the output image.

Note: When decoding to YCbCr, the image will be decoded to RGB and then converted to YCbCr, following the YCbCr definition from ITU-R BT.601.
preallocate_height_hint (int, optional, default = 0) –
Image width hint.

Applies only to the mixed backend type in NVIDIA Ampere GPU architecture.

The hint is used to preallocate memory for the HW JPEG decoder.
preallocate_width_hint (int, optional, default = 0) –
Image width hint.

Applies only to the mixed backend type in NVIDIA Ampere GPU architecture.

The hint is used to preallocate memory for the HW JPEG decoder.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
rel_end (float or list of float or TensorList of float, optional) –
End relative coordinates of the slice (range [0.0 - 1.0].

Note: Providing named arguments start, end, shape, rel_start, rel_end, rel_shape is incompatible with providing positional inputs anchor and shape.
rel_shape (float or list of float or TensorList of float, optional) –
Relative shape of the slice (range [0.0 - 1.0]).

Providing named arguments start, end, shape, rel_start, rel_end, rel_shape is incompatible with providing positional inputs anchor and shape.
rel_start (float or list of float or TensorList of float, optional) –
Start relative coordinates of the slice (range [0.0 - 1.0]).

Note: Providing named arguments start, end, shape, rel_start, rel_end, rel_shape is incompatible with providing positional inputs anchor and shape.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shape (int or list of int or TensorList of int, optional) –
Shape of the slice.

Providing named arguments start, end, shape, rel_start, rel_end, rel_shape is incompatible with providing positional inputs anchor and shape.
split_stages (bool) –

Warning

The argument split_stages is no longer used and will be removed in a future release.
start (int or list of int or TensorList of int, optional) –
Start coordinates of the slice.

Note: Providing named arguments start/end or start/shape is incompatible with providing positional inputs anchor and shape.
use_chunk_allocator (bool) –

Warning

The argument use_chunk_allocator is no longer used and will be removed in a future release.
use_fast_idct (bool, optional, default = False) –
Enables fast IDCT in the libjpeg-turbo based CPU decoder, used when device is set to “cpu” or when the it is set to “mixed” but the particular image can not be handled by the GPU implementation.

According to the libjpeg-turbo documentation, decompression performance is improved by up to 14% with little reduction in quality.

nvidia.dali.fn.experimental¶

nvidia.dali.fn.experimental.readers¶

nvidia.dali.fn.experimental.readers.video(*inputs, **kwargs)¶

Loads and decodes video files using FFmpeg.

The video streams can be in most of the container file formats. FFmpeg is used to parse video containers and returns a batch of sequences of sequence_length frames with shape (N, F, H, W, C), where N is the batch size, and F is the number of frames).

Note

Containers which do not support indexing, like MPEG, require DALI to build the index.

DALI will go through the video and mark keyframes to be able to seek effectively, even in the variable frame rate scenario.

Supported backends

‘cpu’

Keyword Arguments

sequence_length (int) – Frames to load per sequence.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dont_use_mmap (bool, optional, default = False) –
If set to True, the Loader will use plain file I/O instead of trying to map the file in memory.

Mapping provides a small performance benefit when accessing a local file system, but most network file systems, do not provide optimum performance.
filenames (str or list of str, optional, default = []) – Absolute paths to the video files to load.
initial_fill (int, optional, default = 1024) –
Size of the buffer that is used for shuffling.

If random_shuffle is False, this parameter is ignored.
labels (int or list of int, optional) – Labels associated with the files listed in filenames argument. If not provided, no labels will be yielded.
lazy_init (bool, optional, default = False) – Parse and prepare the dataset metadata only during the first run instead of in the constructor.
num_shards (int, optional, default = 1) –
Partitions the data into the specified number of parts (shards).

This is typically used for multi-GPU or multi-node training.
pad_last_batch (bool, optional, default = False) –
If set to True, pads the shard by repeating the last sample.

Note

If the number of batches differs across shards, this option can cause an entire batch of repeated samples to be added to the dataset.
prefetch_queue_depth (int, optional, default = 1) –
Specifies the number of batches to be prefetched by the internal Loader.

This value should be increased when the pipeline is CPU-stage bound, trading memory consumption for better interleaving with the Loader thread.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
random_shuffle (bool, optional, default = False) –
Determines whether to randomly shuffle data.

A prefetch buffer with a size equal to initial_fill is used to read data sequentially, and then samples are selected randomly to form a batch.
read_ahead (bool, optional, default = False) –
Determines whether the accessed data should be read ahead.

For large files such as LMDB, RecordIO, or TFRecord, this argument slows down the first access but decreases the time of all of the following accesses.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shard_id (int, optional, default = 0) – Index of the shard to read.
skip_cached_images (bool, optional, default = False) –
If set to True, the loading data will be skipped when the sample is in the decoder cache.

In this case, the output of the loader will be empty.
step (int, optional, default = -1) –
Frame interval between each sequence.

When the value is less than 0, step is set to sequence_length.
stick_to_shard (bool, optional, default = False) –
Determines whether the reader should stick to a data shard instead of going through the entire dataset.

If decoder caching is used, it significantly reduces the amount of data to be cached, but might affect accuracy of the training.
stride (int, optional, default = 1) – Distance between consecutive frames in the sequence.
tensor_init_bytes (int, optional, default = 1048576) – Hint for how much memory to allocate per image.

nvidia.dali.fn.noise¶

nvidia.dali.fn.noise.gaussian(*inputs, **kwargs)¶

Applies gaussian noise to the input.

The shape and data type of the output will match the input.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
mean (float or TensorList of float, optional, default = 0.0) – Mean of the distribution.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
stddev (float or TensorList of float, optional, default = 1.0) – Standard deviation of the distribution.

nvidia.dali.fn.noise.salt_and_pepper(*inputs, **kwargs)¶

Applies salt-and-pepper noise to the input.

The shape and data type of the output will match the input.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
pepper_val (float or TensorList of float, optional) –
Value of “pepper”.

If not provided, the pepper value will be -1.0 for floating point types or the minimum value of the input data type otherwise, converted to the data type of the input.
per_channel (bool, optional, default = False) –
Determines whether the noise should be generated for each channel independently.

If set to True, the noise is generated for each channel independently, resulting in some channels being corrupted and others kept intact. If set to False, the noise is generated once and applied to all channels, so that all channels in a pixel should either be kept intact, take the “pepper” value, or the “salt” value.

Note: Per-channel noise generation requires the input layout to contain a channels (‘C’) dimension, or be empty. In the case of the layout being empty, channel-last layout is assumed.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
prob (float or TensorList of float, optional, default = 0.05) – Probability of an output value to take a salt or pepper value.
salt_val (float or TensorList of float, optional) –
Value of “salt”.

If not provided, the salt value will be 1.0 for floating point types or the maximum value of the input data type otherwise, converted to the data type of the input.
salt_vs_pepper (float or TensorList of float, optional, default = 0.5) – Probability of a corrupted output value to take a salt value.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.noise.shot(*inputs, **kwargs)¶

Applies shot noise to the input.

The shot noise is generated by applying the following formula:

output[:] = poisson_dist(max(0, input[:] / factor)) * factor) if factor != 0
output[:] = input[:]                                          if factor == 0

where poisson_dist represents a poisson distribution.

Shot noise is a noise that’s present in data generated by a Poisson process, like registering photons by an image sensor. This operator simulates the data acquisition process where each event increases the output value by factor and the input tensor contains the expected values of corresponding output points. For example, a factor of 0.1 means that 10 events are needed to increase the output value by 1, while a factor of 10 means that a single event increases the output by 10. The output values are quantized to multiples of factor. The larger the factor, the more noise is present in the output. A factor of 0 makes this an identity operation.

The shape and data type of the output will match the input.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
factor (float or TensorList of float, optional, default = 20.0) – Factor parameter.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.random¶

nvidia.dali.fn.random.coin_flip(*inputs, **kwargs)¶

Generates random boolean values following a bernoulli distribution.

The probability of generating a value 1 (true) is determined by the probability argument.

The shape of the generated data can be either specified explicitly with a shape argument, or chosen to match the shape of the input, if provided. If none are present, a single value per sample is generated.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList, optional) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional) –
Output data type.

Note

The generated numbers are converted to the output data type, rounding and clamping if necessary.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
probability (float or TensorList of float, optional, default = 0.5) – Probability of value 1.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shape (int or list of int or TensorList of int, optional) – Shape of the output data.

nvidia.dali.fn.random.normal(*inputs, **kwargs)¶

Generates random numbers following a normal distribution.

The shape of the generated data can be either specified explicitly with a shape argument, or chosen to match the shape of the input, if provided. If none are present, a single value per sample is generated.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList, optional) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional) –
Output data type.

Note

The generated numbers are converted to the output data type, rounding and clamping if necessary.
mean (float or TensorList of float, optional, default = 0.0) – Mean of the distribution.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shape (int or list of int or TensorList of int, optional) – Shape of the output data.
stddev (float or TensorList of float, optional, default = 1.0) – Standard deviation of the distribution.

nvidia.dali.fn.random.uniform(*inputs, **kwargs)¶

Generates random numbers following a uniform distribution.

It can be configured to produce a continuous uniform distribution in the range [min, max), or a discrete uniform distribution where any of the specified values [v0, v1, …, vn] occur with equal probability.

The shape of the generated data can be either specified explicitly with a shape argument, or chosen to match the shape of the input, if provided. If none are present, a scalar is generated.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList, optional) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional) –
Output data type.

Note

The generated numbers are converted to the output data type, rounding and clamping if necessary.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
range (float or list of float or TensorList of float, optional, default = [-1.0, 1.0]) –
Range [min, max) of a continuous uniform distribution.

This argument is mutually exclusive with values.

Warning

When specifying an integer type as dtype, the generated numbers can go outside the specified range, due to rounding.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shape (int or list of int or TensorList of int, optional) – Shape of the output data.
values (float or list of float or TensorList of float, optional) –
The discrete values [v0, v1, …, vn] produced by a discrete uniform distribution.

This argument is mutually exclusive with range.

nvidia.dali.fn.readers¶

nvidia.dali.fn.readers.caffe(*inputs, **kwargs)¶

Reads (Image, label) pairs from a Caffe LMDB.

Supported backends

‘cpu’

Keyword Arguments

path (str or list of str) – List of paths to the Caffe LMDB directories.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dont_use_mmap (bool, optional, default = False) –
If set to True, the Loader will use plain file I/O instead of trying to map the file in memory.

Mapping provides a small performance benefit when accessing a local file system, but most network file systems, do not provide optimum performance.
image_available (bool, optional, default = True) – Determines whether an image is available in this LMDB.
initial_fill (int, optional, default = 1024) –
Size of the buffer that is used for shuffling.

If random_shuffle is False, this parameter is ignored.
label_available (bool, optional, default = True) – Determines whether a label is available.
lazy_init (bool, optional, default = False) – Parse and prepare the dataset metadata only during the first run instead of in the constructor.
num_shards (int, optional, default = 1) –
Partitions the data into the specified number of parts (shards).

This is typically used for multi-GPU or multi-node training.
pad_last_batch (bool, optional, default = False) –
If set to True, pads the shard by repeating the last sample.

Note

If the number of batches differs across shards, this option can cause an entire batch of repeated samples to be added to the dataset.
prefetch_queue_depth (int, optional, default = 1) –
Specifies the number of batches to be prefetched by the internal Loader.

This value should be increased when the pipeline is CPU-stage bound, trading memory consumption for better interleaving with the Loader thread.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
random_shuffle (bool, optional, default = False) –
Determines whether to randomly shuffle data.

A prefetch buffer with a size equal to initial_fill is used to read data sequentially, and then samples are selected randomly to form a batch.
read_ahead (bool, optional, default = False) –
Determines whether the accessed data should be read ahead.

For large files such as LMDB, RecordIO, or TFRecord, this argument slows down the first access but decreases the time of all of the following accesses.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shard_id (int, optional, default = 0) – Index of the shard to read.
skip_cached_images (bool, optional, default = False) –
If set to True, the loading data will be skipped when the sample is in the decoder cache.

In this case, the output of the loader will be empty.
stick_to_shard (bool, optional, default = False) –
Determines whether the reader should stick to a data shard instead of going through the entire dataset.

If decoder caching is used, it significantly reduces the amount of data to be cached, but might affect accuracy of the training.
tensor_init_bytes (int, optional, default = 1048576) – Hint for how much memory to allocate per image.

nvidia.dali.fn.readers.caffe2(*inputs, **kwargs)¶

Reads sample data from a Caffe2 Lightning Memory-Mapped Database (LMDB).

Supported backends

‘cpu’

Keyword Arguments

path (str or list of str) – List of paths to the Caffe2 LMDB directories.
additional_inputs (int, optional, default = 0) – Additional auxiliary data tensors that are provided for each sample.
bbox (bool, optional, default = False) – Denotes whether the bounding-box information is present.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dont_use_mmap (bool, optional, default = False) –
If set to True, the Loader will use plain file I/O instead of trying to map the file in memory.

Mapping provides a small performance benefit when accessing a local file system, but most network file systems, do not provide optimum performance.
image_available (bool, optional, default = True) – Determines whether an image is available in this LMDB.
initial_fill (int, optional, default = 1024) –
Size of the buffer that is used for shuffling.

If random_shuffle is False, this parameter is ignored.
label_type (int, optional, default = 0) –
Type of label stored in dataset.

Here is a list of the available values:
- 0 = SINGLE_LABEL: which is the integer label for the multi-class classification.
- 1 = MULTI_LABEL_SPARSE: which is the sparse active label indices for multi-label classification.
- 2 = MULTI_LABEL_DENSE: which is the dense label embedding vector for label embedding regression.
- 3 = MULTI_LABEL_WEIGHTED_SPARSE: which is the sparse active label indices with per-label weights for multi-label classification.
- 4 = NO_LABEL: where no label is available.
lazy_init (bool, optional, default = False) – Parse and prepare the dataset metadata only during the first run instead of in the constructor.
num_labels (int, optional, default = 1) –
Number of classes in the dataset.

Required when sparse labels are used.
num_shards (int, optional, default = 1) –
Partitions the data into the specified number of parts (shards).

This is typically used for multi-GPU or multi-node training.
pad_last_batch (bool, optional, default = False) –
If set to True, pads the shard by repeating the last sample.

Note

If the number of batches differs across shards, this option can cause an entire batch of repeated samples to be added to the dataset.
prefetch_queue_depth (int, optional, default = 1) –
Specifies the number of batches to be prefetched by the internal Loader.

This value should be increased when the pipeline is CPU-stage bound, trading memory consumption for better interleaving with the Loader thread.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
random_shuffle (bool, optional, default = False) –
Determines whether to randomly shuffle data.

A prefetch buffer with a size equal to initial_fill is used to read data sequentially, and then samples are selected randomly to form a batch.
read_ahead (bool, optional, default = False) –
Determines whether the accessed data should be read ahead.

For large files such as LMDB, RecordIO, or TFRecord, this argument slows down the first access but decreases the time of all of the following accesses.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shard_id (int, optional, default = 0) – Index of the shard to read.
skip_cached_images (bool, optional, default = False) –
If set to True, the loading data will be skipped when the sample is in the decoder cache.

In this case, the output of the loader will be empty.
stick_to_shard (bool, optional, default = False) –
Determines whether the reader should stick to a data shard instead of going through the entire dataset.

If decoder caching is used, it significantly reduces the amount of data to be cached, but might affect accuracy of the training.
tensor_init_bytes (int, optional, default = 1048576) – Hint for how much memory to allocate per image.

nvidia.dali.fn.readers.coco(*inputs, **kwargs)¶

Reads data from a COCO dataset that is composed of a directory with images and annotation JSON files.

This readers produces the following outputs:

images, bounding_boxes, labels, ((polygons, vertices) | (pixelwise_masks)), (image_ids)

images Each sample contains image data with layout HWC (height, width, channels).
bounding_boxes Each sample can have an arbitrary M number of bounding boxes, each described by 4 coordinates:
```
[[x_0, y_0, w_0, h_0],
 [x_1, y_1, w_1, h_1]
 ...
 [x_M, y_M, w_M, h_M]]
```
or in [l, t, r, b] format if requested (see ltrb argument).
labels Each bounding box is associated with an integer label representing a category identifier:
```
[label_0, label_1, ..., label_M]
```
polygons and vertices (Optional, present if polygon_masks is set to True) If polygon_masks is enabled, two extra outputs describing masks by a set of polygons. Each mask contains an arbitrary number of polygons P, each associated with a mask index in the range [0, M) and composed by a group of V vertices. The output polygons describes the polygons as follows:
```
[[mask_idx_0, start_vertex_idx_0, end_vertex_idx_0],
 [mask_idx_1, start_vertex_idx_1, end_vertex_idx_1],
 ...
 [mask_idx_P, start_vertex_idx_P, end_vertex_idx_P]]
```
where mask_idx is the index of the mask the polygon, in the range [0, M), and start_vertex_idx and end_verted_idx define the range of indices of vertices, as they appear in the output vertices, belonging to this polygon. Each sample in vertices contains a list of vertices that composed the different polygons in the sample, as 2D coordinates:
```
[[x_0, y_0],
 [x_1, y_1],
 ...
 [x_V, y_V]]
```
pixelwise_masks (Optional, present if argument pixelwise_masks is set to True) Contains image-like data, same shape and layout as images, representing a pixelwise segmentation mask.
image_ids (Optional, present if argument image_ids is set to True) One element per sample, representing an image identifier.

Supported backends

‘cpu’

Keyword Arguments

annotations_file (str, optional, default = ‘’) – List of paths to the JSON annotations files.
avoid_class_remapping (bool, optional, default = False) –
If set to True, lasses ID values are returned directly as they are defined in the manifest file.

Otherwise, classes’ ID values are mapped to consecutive values in range 1-number of classes disregarding exact values from the manifest (0 is reserved for a special background class.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dont_use_mmap (bool, optional, default = False) –
If set to True, the Loader will use plain file I/O instead of trying to map the file in memory.

Mapping provides a small performance benefit when accessing a local file system, but most network file systems, do not provide optimum performance.
file_root (str, optional) –
Path to a directory that contains the data files.

If a file list is not provided, this argument is required.
image_ids (bool, optional, default = False) – If set to True, the image IDs will be produced in an extra output.
images (str or list of str, optional) –
A list of image paths.

If provided, it specifies the images that will be read. The images will be read in the same order as they appear in the list, and in case of duplicates, multiple copies of the relevant samples will be produced.

If left unspecified or set to None, all images listed in the annotation file are read exactly once, ordered by their image id.

The paths to be kept should match exactly those in the annotations file.

Note: This argument is mutually exclusive with preprocessed_annotations.
initial_fill (int, optional, default = 1024) –
Size of the buffer that is used for shuffling.

If random_shuffle is False, this parameter is ignored.
lazy_init (bool, optional, default = False) – Parse and prepare the dataset metadata only during the first run instead of in the constructor.
ltrb (bool, optional, default = False) –
If set to True, bboxes are returned as [left, top, right, bottom].

If set to False, the bboxes are returned as [x, y, width, height].
masks (bool, optional, default = False) –
Enable polygon masks.

Warning

Use polygon_masks instead. Note that the polygon format has changed mask_id, start_coord, end_coord to mask_id, start_vertex, end_vertex where start_coord and end_coord are total number of coordinates, effectly start_coord = 2 * start_vertex and end_coord = 2 * end_vertex. Example: A polygon with vertices [[x0, y0], [x1, y1], [x2, y2]] would be represented as [mask_id, 0, 6] when using the deprecated argument masks, but [mask_id, 0, 3] when using the new argument polygon_masks.
num_shards (int, optional, default = 1) –
Partitions the data into the specified number of parts (shards).

This is typically used for multi-GPU or multi-node training.
pad_last_batch (bool, optional, default = False) –
If set to True, pads the shard by repeating the last sample.

Note

If the number of batches differs across shards, this option can cause an entire batch of repeated samples to be added to the dataset.
pixelwise_masks (bool, optional, default = False) – If true, segmentation masks are read and returned as pixel-wise masks. This argument is mutually exclusive with polygon_masks.
polygon_masks (bool, optional, default = False) –
If set to True, segmentation mask polygons are read in the form of two outputs: polygons and vertices. This argument is mutually exclusive with pixelwise_masks.

Warning

Currently objects with iscrowd=1 annotations are skipped.
prefetch_queue_depth (int, optional, default = 1) –
Specifies the number of batches to be prefetched by the internal Loader.

This value should be increased when the pipeline is CPU-stage bound, trading memory consumption for better interleaving with the Loader thread.
preprocessed_annotations (str, optional, default = ‘’) – Path to the directory with meta files that contain preprocessed COCO annotations.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
random_shuffle (bool, optional, default = False) –
Determines whether to randomly shuffle data.

A prefetch buffer with a size equal to initial_fill is used to read data sequentially, and then samples are selected randomly to form a batch.
ratio (bool, optional, default = False) – If set to True, the returned bbox and mask polygon coordinates are relative to the image dimensions.
read_ahead (bool, optional, default = False) –
Determines whether the accessed data should be read ahead.

For large files such as LMDB, RecordIO, or TFRecord, this argument slows down the first access but decreases the time of all of the following accesses.
save_preprocessed_annotations (bool, optional, default = False) – If set to True, the operator saves a set of files containing binary representations of the preprocessed COCO annotations.
save_preprocessed_annotations_dir (str, optional, default = ‘’) – Path to the directory in which to save the preprocessed COCO annotations files.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shard_id (int, optional, default = 0) – Index of the shard to read.
shuffle_after_epoch (bool, optional, default = False) – If set to True, the reader shuffles the entire dataset after each epoch.
size_threshold (float, optional, default = 0.1) – If the width or the height, in number of pixels, of a bounding box that represents an instance of an object is lower than this value, the object will be ignored.
skip_cached_images (bool, optional, default = False) –
If set to True, the loading data will be skipped when the sample is in the decoder cache.

In this case, the output of the loader will be empty.
skip_empty (bool, optional, default = False) – If true, reader will skip samples with no object instances in them
stick_to_shard (bool, optional, default = False) –
Determines whether the reader should stick to a data shard instead of going through the entire dataset.

If decoder caching is used, it significantly reduces the amount of data to be cached, but might affect accuracy of the training.
tensor_init_bytes (int, optional, default = 1048576) – Hint for how much memory to allocate per image.
dump_meta_files (bool) –

Warning

The argument dump_meta_files is a deprecated alias for save_preprocessed_annotations. Use save_preprocessed_annotations instead.
dump_meta_files_path (str) –

Warning

The argument dump_meta_files_path is a deprecated alias for save_preprocessed_annotations_dir. Use save_preprocessed_annotations_dir instead.
meta_files_path (str) –

Warning

The argument meta_files_path is a deprecated alias for preprocessed_annotations. Use preprocessed_annotations instead.
save_img_ids (bool) –

Warning

The argument save_img_ids is a deprecated alias for image_ids. Use image_ids instead.

nvidia.dali.fn.readers.file(*inputs, **kwargs)¶

Reads file contents and returns file-label pairs.

This operator can be used in the following modes:

Listing files from a directory, assigning labels based on subdirectory structure.

In this mode, the directory indicated in file_root argument should contain one or more subdirectories. The files in these subdirectories are listed and assigned labels based on lexicographical order of the subdirectory. If you provide file_filters argument with a list of glob strings, the operator will list files matching at least one of the patterns. Otherwise, a default set of filters is used (see the default value of file_filters for details).

For example, this directory structure:

<file_root>/0/image0.jpg
<file_root>/0/world_map.jpg
<file_root>/0/antarctic.png
<file_root>/1/cat.jpeg
<file_root>/1/dog.tif
<file_root>/2/car.jpeg
<file_root>/2/truck.jp2

by default will yield the following outputs:

<contents of 0/image0.jpg>        0
<contents of 0/world_map.jpg>     0
<contents of 0/antarctic.png>     0
<contents of 1/cat.jpeg>          1
<contents of 1/dog.tif>           1
<contents of 2/car.jpeg>          2
<contents of 2/truck.jp2>         2

and with file_filters = ["*.jpg", "*.jpeg"] will yield the following outputs:

<contents of 0/image0.jpg>        0
<contents of 0/world_map.jpg>     0
<contents of 1/cat.jpeg>          1
<contents of 2/car.jpeg>          2

Use file names and labels stored in a text file.

file_list argument points to a file which contains one file name and label per line. Example:

dog.jpg 0
cute kitten.jpg 1
doge.png 0

The file names can contain spaces in the middle, but cannot contain trailing whitespace.

Use file names and labels provided as a list of strings and integers, respectively.

As with other readers, the (file, label) pairs returned by this operator can be randomly shuffled and various sharding strategies can be applied. See documentation of this operator’s arguments for details.

Supported backends

‘cpu’

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
case_sensitive_filter (bool, optional, default = False) – If set to True, the filter will be matched case-sensitively, otherwise case-insensitively.
dont_use_mmap (bool, optional, default = False) –
If set to True, the Loader will use plain file I/O instead of trying to map the file in memory.

Mapping provides a small performance benefit when accessing a local file system, but most network file systems, do not provide optimum performance.
file_filters (str or list of str, optional, default = [‘*.jpg’, ‘*.jpeg’, ‘*.png’, ‘*.bmp’, ‘*.tif’, ‘*.tiff’, ‘*.pnm’, ‘*.ppm’, ‘*.pgm’, ‘*.pbm’, ‘*.jp2’, ‘*.webp’, ‘*.flac’, ‘*.ogg’, ‘*.wav’]) –
A list of glob strings to filter the list of files in the sub-directories of the file_root.

This argument is ignored when file paths are taken from file_list or files.
file_list (str, optional) –
Path to a text file that contains one whitespace-separated filename label pair per line. The filenames are relative to the location of that file or to file_root, if specified.

This argument is mutually exclusive with files.
file_root (str, optional) –
Path to a directory that contains the data files.

If not using file_list or files, this directory is traversed to discover the files. file_root is required in this mode of operation.
files (str or list of str, optional) –
A list of file paths to read the data from.

If file_root is provided, the paths are treated as being relative to it. When using files, the labels are taken from labels argument or, if it was not supplied, contain indices at which given file appeared in the files list.

This argument is mutually exclusive with file_list.
initial_fill (int, optional, default = 1024) –
Size of the buffer that is used for shuffling.

If random_shuffle is False, this parameter is ignored.
labels (int or list of int, optional) –
Labels accompanying contents of files listed in files argument.

If not used, sequential 0-based indices are used as labels
lazy_init (bool, optional, default = False) – Parse and prepare the dataset metadata only during the first run instead of in the constructor.
num_shards (int, optional, default = 1) –
Partitions the data into the specified number of parts (shards).

This is typically used for multi-GPU or multi-node training.
pad_last_batch (bool, optional, default = False) –
If set to True, pads the shard by repeating the last sample.

Note

If the number of batches differs across shards, this option can cause an entire batch of repeated samples to be added to the dataset.
prefetch_queue_depth (int, optional, default = 1) –
Specifies the number of batches to be prefetched by the internal Loader.

This value should be increased when the pipeline is CPU-stage bound, trading memory consumption for better interleaving with the Loader thread.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
random_shuffle (bool, optional, default = False) –
Determines whether to randomly shuffle data.

A prefetch buffer with a size equal to initial_fill is used to read data sequentially, and then samples are selected randomly to form a batch.
read_ahead (bool, optional, default = False) –
Determines whether the accessed data should be read ahead.

For large files such as LMDB, RecordIO, or TFRecord, this argument slows down the first access but decreases the time of all of the following accesses.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shard_id (int, optional, default = 0) – Index of the shard to read.
shuffle_after_epoch (bool, optional, default = False) –
If set to True, the reader shuffles the entire dataset after each epoch.

stick_to_shard and random_shuffle cannot be used when this argument is set to True.
skip_cached_images (bool, optional, default = False) –
If set to True, the loading data will be skipped when the sample is in the decoder cache.

In this case, the output of the loader will be empty.
stick_to_shard (bool, optional, default = False) –
Determines whether the reader should stick to a data shard instead of going through the entire dataset.

If decoder caching is used, it significantly reduces the amount of data to be cached, but might affect accuracy of the training.
tensor_init_bytes (int, optional, default = 1048576) – Hint for how much memory to allocate per image.

nvidia.dali.fn.readers.mxnet(*inputs, **kwargs)¶

Reads the data from an MXNet RecordIO.

Supported backends

‘cpu’

Keyword Arguments

index_path (str or list of str) –
List (of length 1) that contains a path to the index (.idx) file.

The file is generated by the MXNet’s im2rec.py script with the RecordIO file. The list can also be generated by using the rec2idx script that is distributed with DALI.
path (str or list of str) – List of paths to the RecordIO files.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dont_use_mmap (bool, optional, default = False) –
If set to True, the Loader will use plain file I/O instead of trying to map the file in memory.

Mapping provides a small performance benefit when accessing a local file system, but most network file systems, do not provide optimum performance.
initial_fill (int, optional, default = 1024) –
Size of the buffer that is used for shuffling.

If random_shuffle is False, this parameter is ignored.
lazy_init (bool, optional, default = False) – Parse and prepare the dataset metadata only during the first run instead of in the constructor.
num_shards (int, optional, default = 1) –
Partitions the data into the specified number of parts (shards).

This is typically used for multi-GPU or multi-node training.
pad_last_batch (bool, optional, default = False) –
If set to True, pads the shard by repeating the last sample.

Note

If the number of batches differs across shards, this option can cause an entire batch of repeated samples to be added to the dataset.
prefetch_queue_depth (int, optional, default = 1) –
Specifies the number of batches to be prefetched by the internal Loader.

This value should be increased when the pipeline is CPU-stage bound, trading memory consumption for better interleaving with the Loader thread.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
random_shuffle (bool, optional, default = False) –
Determines whether to randomly shuffle data.

A prefetch buffer with a size equal to initial_fill is used to read data sequentially, and then samples are selected randomly to form a batch.
read_ahead (bool, optional, default = False) –
Determines whether the accessed data should be read ahead.

For large files such as LMDB, RecordIO, or TFRecord, this argument slows down the first access but decreases the time of all of the following accesses.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shard_id (int, optional, default = 0) – Index of the shard to read.
skip_cached_images (bool, optional, default = False) –
If set to True, the loading data will be skipped when the sample is in the decoder cache.

In this case, the output of the loader will be empty.
stick_to_shard (bool, optional, default = False) –
Determines whether the reader should stick to a data shard instead of going through the entire dataset.

If decoder caching is used, it significantly reduces the amount of data to be cached, but might affect accuracy of the training.
tensor_init_bytes (int, optional, default = 1048576) – Hint for how much memory to allocate per image.

nvidia.dali.fn.readers.nemo_asr(*inputs, **kwargs)¶

Reads automatic speech recognition (ASR) data (audio, text) from an NVIDIA NeMo compatible manifest.

Example manifest file:

{"audio_filepath": "path/to/audio1.wav", "duration": 3.45, "text": "this is a nemo tutorial"}
{"audio_filepath": "path/to/audio1.wav", "offset": 3.45, "duration": 1.45, "text": "same audio file but using offset"}
{"audio_filepath": "path/to/audio2.wav", "duration": 3.45, "text": "third transcript in this example"}

Note

Only audio_filepath is field mandatory. If duration is not specified, the whole audio file will be used. A missing text field will produce an empty string as a text.

Warning

Handling of duration and offset fields is not yet implemented. The current implementation always reads the whole audio file.

This reader produces between 1 and 3 outputs:

Decoded audio data: float, shape=(audio_length,)
(optional, if read_sample_rate=True) Audio sample rate: float, shape=(1,)
(optional, if read_text=True) Transcript text as a null terminated string: uint8, shape=(text_len + 1,)
(optional, if read_idxs=True) Index of the manifest entry: int64, shape=(1,)

Supported backends

‘cpu’

Keyword Arguments

manifest_filepaths (str or list of str) – List of paths to NeMo’s compatible manifest files.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dont_use_mmap (bool, optional, default = False) –
If set to True, the Loader will use plain file I/O instead of trying to map the file in memory.

Mapping provides a small performance benefit when accessing a local file system, but most network file systems, do not provide optimum performance.
downmix (bool, optional, default = True) – If True, downmix all input channels to mono. If downmixing is turned on, decoder will produce always 1-D output
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.FLOAT) –
Output data type.

Supported types: INT16, INT32, and FLOAT.
initial_fill (int, optional, default = 1024) –
Size of the buffer that is used for shuffling.

If random_shuffle is False, this parameter is ignored.
lazy_init (bool, optional, default = False) – Parse and prepare the dataset metadata only during the first run instead of in the constructor.
max_duration (float, optional, default = 0.0) –
If a value greater than 0 is provided, it specifies the maximum allowed duration, in seconds, of the audio samples.

Samples with a duration longer than this value will be ignored.
min_duration (float, optional, default = 0.0) –

If a value greater than 0 is provided, it specifies the minimum allowed duration,
in seconds, of the audio samples.

Samples with a duration shorter than this value will be ignored.
normalize_text (bool) –

Warning

The argument normalize_text is no longer used and will be removed in a future release.
num_shards (int, optional, default = 1) –
Partitions the data into the specified number of parts (shards).

This is typically used for multi-GPU or multi-node training.
pad_last_batch (bool, optional, default = False) –
If set to True, pads the shard by repeating the last sample.

Note

If the number of batches differs across shards, this option can cause an entire batch of repeated samples to be added to the dataset.
prefetch_queue_depth (int, optional, default = 1) –
Specifies the number of batches to be prefetched by the internal Loader.

This value should be increased when the pipeline is CPU-stage bound, trading memory consumption for better interleaving with the Loader thread.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
quality (float, optional, default = 50.0) –
Resampling quality, 0 is lowest, 100 is highest.

0 corresponds to 3 lobes of the sinc filter; 50 gives 16 lobes and 100 gives 64 lobes.
random_shuffle (bool, optional, default = False) –
Determines whether to randomly shuffle data.

A prefetch buffer with a size equal to initial_fill is used to read data sequentially, and then samples are selected randomly to form a batch.
read_ahead (bool, optional, default = False) –
Determines whether the accessed data should be read ahead.

For large files such as LMDB, RecordIO, or TFRecord, this argument slows down the first access but decreases the time of all of the following accesses.
read_idxs (bool, optional, default = False) –

Whether to output the indices of samples as they occur in the manifest file
as a separate output
read_sample_rate (bool, optional, default = True) – Whether to output the sample rate for each sample as a separate output
read_text (bool, optional, default = True) – Whether to output the transcript text for each sample as a separate output
sample_rate (float, optional, default = -1.0) – If specified, the target sample rate, in Hz, to which the audio is resampled.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shard_id (int, optional, default = 0) – Index of the shard to read.
shuffle_after_epoch (bool, optional, default = False) – If true, reader shuffles whole dataset after each epoch
skip_cached_images (bool, optional, default = False) –
If set to True, the loading data will be skipped when the sample is in the decoder cache.

In this case, the output of the loader will be empty.
stick_to_shard (bool, optional, default = False) –
Determines whether the reader should stick to a data shard instead of going through the entire dataset.

If decoder caching is used, it significantly reduces the amount of data to be cached, but might affect accuracy of the training.
tensor_init_bytes (int, optional, default = 1048576) – Hint for how much memory to allocate per image.

nvidia.dali.fn.readers.numpy(*inputs, **kwargs)¶

Reads Numpy arrays from a directory.

This operator can be used in the following modes:

Read all files from a directory indicated by file_root that match given file_filter.
Read file names from a text file indicated in file_list argument.
Read files listed in files argument.

Note

The gpu backend requires cuFile/GDS support (418.x driver family or newer). Please check the relevant GDS package for more details.

Supported backends

‘cpu’
‘gpu’

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
cache_header_information (bool, optional, default = False) – If set to True, the header information for each file is cached, improving access speed.
dont_use_mmap (bool, optional, default = False) –
If set to True, the Loader will use plain file I/O instead of trying to map the file in memory.

Mapping provides a small performance benefit when accessing a local file system, but most network file systems, do not provide optimum performance.
file_filter (str, optional, default = ‘*.npy’) –
If a value is specified, the string is interpreted as glob string to filter the list of files in the sub-directories of the file_root.

This argument is ignored when file paths are taken from file_list or files.
file_list (str, optional) –
Path to a text file that contains filenames (one per line) where the filenames are relative to the location of that file or to file_root, if specified.

This argument is mutually exclusive with files.
file_root (str, optional) –
Path to a directory that contains the data files.

If not using file_list or files, this directory is traversed to discover the files. file_root is required in this mode of operation.
files (str or list of str, optional) –
A list of file paths to read the data from.

If file_root is provided, the paths are treated as being relative to it.

This argument is mutually exclusive with file_list.
fill_value (float, optional, default = 0.0) – Determines the padding value when out_of_bounds_policy is set to “pad”.
initial_fill (int, optional, default = 1024) –
Size of the buffer that is used for shuffling.

If random_shuffle is False, this parameter is ignored.
lazy_init (bool, optional, default = False) – Parse and prepare the dataset metadata only during the first run instead of in the constructor.
num_shards (int, optional, default = 1) –
Partitions the data into the specified number of parts (shards).

This is typically used for multi-GPU or multi-node training.
out_of_bounds_policy (str, optional, default = ‘error’) –
Determines the policy when reading outside of the bounds of the numpy array.

Here is a list of the supported values:
- "error" (default): Attempting to read outside of the bounds of the image will produce an error.
- "pad": The array will be padded as needed with zeros or any other value that is specified with the fill_value argument.
- "trim_to_shape": The ROI will be cut to the bounds of the array.
pad_last_batch (bool, optional, default = False) –
If set to True, pads the shard by repeating the last sample.

Note

If the number of batches differs across shards, this option can cause an entire batch of repeated samples to be added to the dataset.
prefetch_queue_depth (int, optional, default = 1) –
Specifies the number of batches to be prefetched by the internal Loader.

This value should be increased when the pipeline is CPU-stage bound, trading memory consumption for better interleaving with the Loader thread.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
random_shuffle (bool, optional, default = False) –
Determines whether to randomly shuffle data.

A prefetch buffer with a size equal to initial_fill is used to read data sequentially, and then samples are selected randomly to form a batch.
read_ahead (bool, optional, default = False) –
Determines whether the accessed data should be read ahead.

For large files such as LMDB, RecordIO, or TFRecord, this argument slows down the first access but decreases the time of all of the following accesses.
register_buffers (bool, optional, default = True) –
Applies only to the gpu backend type.

Warning

This argument is temporarily disabled and left for backward compatibility. It will be reenabled in the future releases.

If true, the device I/O buffers will be registered with cuFile. It is not recommended if the sample sizes vary a lot.
rel_roi_end (float or list of float or TensorList of float, optional) –
End of the region-of-interest, in relative coordinates (range [0.0 - 1.0]).

This argument is incompatible with “roi_end”, “roi_shape” and “rel_roi_shape”.
rel_roi_shape (float or list of float or TensorList of float, optional) –
Shape of the region-of-interest, in relative coordinates (range [0.0 - 1.0]).

This argument is incompatible with “roi_shape”, “roi_end” and “rel_roi_end”.
rel_roi_start (float or list of float or TensorList of float, optional) –
Start of the region-of-interest, in relative coordinates (range [0.0 - 1.0]).

This argument is incompatible with “roi_start”.
roi_axes (int or list of int, optional, default = []) –
Order of dimensions used for the ROI anchor and shape argumens, as dimension indices.

If not provided, all the dimensions should be specified in the ROI arguments.
roi_end (int or list of int or TensorList of int, optional) –
End of the region-of-interest, in absolute coordinates.

This argument is incompatible with “rel_roi_end”, “roi_shape” and “rel_roi_shape”.
roi_shape (int or list of int or TensorList of int, optional) –
Shape of the region-of-interest, in absolute coordinates.

This argument is incompatible with “rel_roi_shape”, “roi_end” and “rel_roi_end”.
roi_start (int or list of int or TensorList of int, optional) –
Start of the region-of-interest, in absolute coordinates.

This argument is incompatible with “rel_roi_start”.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shard_id (int, optional, default = 0) – Index of the shard to read.
shuffle_after_epoch (bool, optional, default = False) –
If set to True, the reader shuffles the entire dataset after each epoch.

stick_to_shard and random_shuffle cannot be used when this argument is set to True.
skip_cached_images (bool, optional, default = False) –
If set to True, the loading data will be skipped when the sample is in the decoder cache.

In this case, the output of the loader will be empty.
stick_to_shard (bool, optional, default = False) –
Determines whether the reader should stick to a data shard instead of going through the entire dataset.

If decoder caching is used, it significantly reduces the amount of data to be cached, but might affect accuracy of the training.
tensor_init_bytes (int, optional, default = 1048576) – Hint for how much memory to allocate per image.

nvidia.dali.fn.readers.sequence(*inputs, **kwargs)¶

Reads [Frame] sequences from a directory representing a collection of streams.

This operator expects file_root to contain a set of directories, where each directory represents an extracted video stream. This stream is represented by one file for each frame, sorted lexicographically. Sequences do not cross the stream boundary and only complete sequences are considered, so there is no padding.

Example directory structure:

- file_root
  - 0
    - 00001.png
    - 00002.png
    - 00003.png
    - 00004.png
    - 00005.png
    - 00006.png
    ....

  - 1
    - 00001.png
    - 00002.png
    - 00003.png
    - 00004.png
    - 00005.png
    - 00006.png
    ....

Note

This operator is an analogue of VideoReader working on video frames extracted as separate images. It’s main purpose is for test baseline. For regular usage, the VideoReader is the recommended approach.

This operator allows sequence inputs.

Supported backends

‘cpu’

Keyword Arguments

file_root (str) – Path to a directory containing streams, where the directories represent streams.
sequence_length (int) – Length of sequence to load for each sample.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dont_use_mmap (bool, optional, default = False) –
If set to True, the Loader will use plain file I/O instead of trying to map the file in memory.

Mapping provides a small performance benefit when accessing a local file system, but most network file systems, do not provide optimum performance.
image_type (nvidia.dali.types.DALIImageType, optional, default = DALIImageType.RGB) – The color space of input and output image.
initial_fill (int, optional, default = 1024) –
Size of the buffer that is used for shuffling.

If random_shuffle is False, this parameter is ignored.
lazy_init (bool, optional, default = False) – Parse and prepare the dataset metadata only during the first run instead of in the constructor.
num_shards (int, optional, default = 1) –
Partitions the data into the specified number of parts (shards).

This is typically used for multi-GPU or multi-node training.
pad_last_batch (bool, optional, default = False) –
If set to True, pads the shard by repeating the last sample.

Note

If the number of batches differs across shards, this option can cause an entire batch of repeated samples to be added to the dataset.
prefetch_queue_depth (int, optional, default = 1) –
Specifies the number of batches to be prefetched by the internal Loader.

This value should be increased when the pipeline is CPU-stage bound, trading memory consumption for better interleaving with the Loader thread.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
random_shuffle (bool, optional, default = False) –
Determines whether to randomly shuffle data.

A prefetch buffer with a size equal to initial_fill is used to read data sequentially, and then samples are selected randomly to form a batch.
read_ahead (bool, optional, default = False) –
Determines whether the accessed data should be read ahead.

For large files such as LMDB, RecordIO, or TFRecord, this argument slows down the first access but decreases the time of all of the following accesses.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shard_id (int, optional, default = 0) – Index of the shard to read.
skip_cached_images (bool, optional, default = False) –
If set to True, the loading data will be skipped when the sample is in the decoder cache.

In this case, the output of the loader will be empty.
step (int, optional, default = 1) – Distance between first frames of consecutive sequences.
stick_to_shard (bool, optional, default = False) –
Determines whether the reader should stick to a data shard instead of going through the entire dataset.

If decoder caching is used, it significantly reduces the amount of data to be cached, but might affect accuracy of the training.
stride (int, optional, default = 1) – Distance between consecutive frames in a sequence.
tensor_init_bytes (int, optional, default = 1048576) – Hint for how much memory to allocate per image.

nvidia.dali.fn.readers.tfrecord(*inputs, **kwargs)¶

Reads samples from a TensorFlow TFRecord file.

Supported backends

‘cpu’

Keyword Arguments

features (dict of (string, nvidia.dali.tfrecord.Feature)) –
A dictionary that maps names of the TFRecord features to extract to the feature type.

Typically obtained by using the dali.tfrecord.FixedLenFeature and dali.tfrecord.VarLenFeature helper functions, which are equal to TensorFlow’s tf.FixedLenFeature and tf.VarLenFeature types, respectively. For additional flexibility, dali.tfrecord.VarLenFeature supports the partial_shape parameter. If provided, the data will be reshaped to match its value, and the first dimension will be inferred from the data size.

If the named feature doesn’t exists in the processed TFRecord entry an empty tensor is returned.
index_path (str or list of str) –
List of paths to index files. There should be one index file for every TFRecord file.

The index files can be obtained from TFRecord files by using the tfrecord2idx script that is distributed with DALI.
path (str or list of str) – List of paths to TFRecord files.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dont_use_mmap (bool, optional, default = False) –
If set to True, the Loader will use plain file I/O instead of trying to map the file in memory.

Mapping provides a small performance benefit when accessing a local file system, but most network file systems, do not provide optimum performance.
initial_fill (int, optional, default = 1024) –
Size of the buffer that is used for shuffling.

If random_shuffle is False, this parameter is ignored.
lazy_init (bool, optional, default = False) – Parse and prepare the dataset metadata only during the first run instead of in the constructor.
num_shards (int, optional, default = 1) –
Partitions the data into the specified number of parts (shards).

This is typically used for multi-GPU or multi-node training.
pad_last_batch (bool, optional, default = False) –
If set to True, pads the shard by repeating the last sample.

Note

If the number of batches differs across shards, this option can cause an entire batch of repeated samples to be added to the dataset.
prefetch_queue_depth (int, optional, default = 1) –
Specifies the number of batches to be prefetched by the internal Loader.

This value should be increased when the pipeline is CPU-stage bound, trading memory consumption for better interleaving with the Loader thread.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
random_shuffle (bool, optional, default = False) –
Determines whether to randomly shuffle data.

A prefetch buffer with a size equal to initial_fill is used to read data sequentially, and then samples are selected randomly to form a batch.
read_ahead (bool, optional, default = False) –
Determines whether the accessed data should be read ahead.

For large files such as LMDB, RecordIO, or TFRecord, this argument slows down the first access but decreases the time of all of the following accesses.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shard_id (int, optional, default = 0) – Index of the shard to read.
skip_cached_images (bool, optional, default = False) –
If set to True, the loading data will be skipped when the sample is in the decoder cache.

In this case, the output of the loader will be empty.
stick_to_shard (bool, optional, default = False) –
Determines whether the reader should stick to a data shard instead of going through the entire dataset.

If decoder caching is used, it significantly reduces the amount of data to be cached, but might affect accuracy of the training.
tensor_init_bytes (int, optional, default = 1048576) – Hint for how much memory to allocate per image.

nvidia.dali.fn.readers.video(*inputs, **kwargs)¶

Loads and decodes video files using FFmpeg and NVDECODE, which is the hardware-accelerated video decoding feature in the NVIDIA(R) GPU.

The video streams can be in most of the container file formats. FFmpeg is used to parse video containers and returns a batch of sequences of sequence_length frames with shape (N, F, H, W, C), where N is the batch size, and F is the number of frames). This class only supports constant frame rate videos.

Note

Containers which doesn’t support indexing, like mpeg, requires DALI to seek to the sequence when each new sequence needs to be decoded.

Supported backends

‘gpu’

Keyword Arguments

sequence_length (int) – Frames to load per sequence.
additional_decode_surfaces (int, optional, default = 2) –
Additional decode surfaces to use beyond minimum required.

This argument is ignored when the decoder cannot determine the minimum number of decode surfaces

Note

This can happen when the driver is an older version.

This parameter can be used to trade off memory usage with performance.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
channels (int, optional, default = 3) – Number of channels.
dont_use_mmap (bool, optional, default = False) –
If set to True, the Loader will use plain file I/O instead of trying to map the file in memory.

Mapping provides a small performance benefit when accessing a local file system, but most network file systems, do not provide optimum performance.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.UINT8) –
Output data type.

Supported types: UINT8 or FLOAT.
enable_frame_num (bool, optional, default = False) – If the file_list or filenames argument is passed, returns the frame number output.
enable_timestamps (bool, optional, default = False) – If the file_list or filenames argument is passed, returns the timestamps output.
file_list (str, optional, default = ‘’) –
Path to the file with a list of file label [start_frame [end_frame]] values.

Positive value means the exact frame, negative counts as a Nth frame from the end (it follows python array indexing schema), equal values for the start and end frame would yield an empty sequence and a warning. This option is mutually exclusive with filenames and file_root.
file_list_frame_num (bool, optional, default = False) –
If the start/end timestamps are provided in file_list, you can interpret them as frame numbers instead of as timestamps.

If floating point values have been provided, the start frame number will be rounded up and the end frame number will be rounded down.

Frame numbers start from 0.
file_list_include_preceding_frame (bool, optional, default = False) –
Changes the behavior how file_list start and end frame timestamps are translated to a frame number.

If the start/end timestamps are provided in file_list as timestamps, the start frame is calculated as ceil(start_time_stamp * FPS) and the end as floor(end_time_stamp * FPS). If this argument is set to True, the equation changes to floor(start_time_stamp * FPS) and ceil(end_time_stamp * FPS) respectively. In effect, the first returned frame is not later, and the end frame not earlier, than the provided timestamps. This behavior is more aligned with how the visible timestamps are correlated with displayed video frames.

Note

When file_list_frame_num is set to True, this option does not take any effect.

Warning

This option is available for legacy behavior compatibility.
file_root (str, optional, default = ‘’) –
Path to a directory that contains the data files.

This option is mutually exclusive with filenames and file_list.
filenames (str or list of str, optional, default = []) –
File names of the video files to load.

This option is mutually exclusive with file_list and file_root.
image_type (nvidia.dali.types.DALIImageType, optional, default = DALIImageType.RGB) – The color space of the output frames (RGB or YCbCr).
initial_fill (int, optional, default = 1024) –
Size of the buffer that is used for shuffling.

If random_shuffle is False, this parameter is ignored.
labels (int or list of int, optional) –
Labels associated with the files listed in filenames argument.

If an empty list is provided, sequential 0-based indices are used as labels. If not provided, no labels will be yielded.
lazy_init (bool, optional, default = False) – Parse and prepare the dataset metadata only during the first run instead of in the constructor.
normalized (bool, optional, default = False) – Gets the output as normalized data.
num_shards (int, optional, default = 1) –
Partitions the data into the specified number of parts (shards).

This is typically used for multi-GPU or multi-node training.
pad_last_batch (bool, optional, default = False) –
If set to True, pads the shard by repeating the last sample.

Note

If the number of batches differs across shards, this option can cause an entire batch of repeated samples to be added to the dataset.
pad_sequences (bool, optional, default = False) –
Allows creation of incomplete sequences if there is an insufficient number of frames at the very end of the video.

Redundant frames are zeroed. Corresponding time stamps and frame numbers are set to -1.
prefetch_queue_depth (int, optional, default = 1) –
Specifies the number of batches to be prefetched by the internal Loader.

This value should be increased when the pipeline is CPU-stage bound, trading memory consumption for better interleaving with the Loader thread.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
random_shuffle (bool, optional, default = False) –
Determines whether to randomly shuffle data.

A prefetch buffer with a size equal to initial_fill is used to read data sequentially, and then samples are selected randomly to form a batch.
read_ahead (bool, optional, default = False) –
Determines whether the accessed data should be read ahead.

For large files such as LMDB, RecordIO, or TFRecord, this argument slows down the first access but decreases the time of all of the following accesses.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shard_id (int, optional, default = 0) – Index of the shard to read.
skip_cached_images (bool, optional, default = False) –
If set to True, the loading data will be skipped when the sample is in the decoder cache.

In this case, the output of the loader will be empty.
skip_vfr_check (bool, optional, default = False) –
Skips the check for the variable frame rate (VFR) videos.

Use this flag to suppress false positive detection of VFR videos.

Warning

When the dataset indeed contains VFR files, setting this flag may cause the decoder to malfunction.
step (int, optional, default = -1) –
Frame interval between each sequence.

When the value is less than 0, step is set to sequence_length.
stick_to_shard (bool, optional, default = False) –
Determines whether the reader should stick to a data shard instead of going through the entire dataset.

If decoder caching is used, it significantly reduces the amount of data to be cached, but might affect accuracy of the training.
stride (int, optional, default = 1) – Distance between consecutive frames in the sequence.
tensor_init_bytes (int, optional, default = 1048576) – Hint for how much memory to allocate per image.

nvidia.dali.fn.readers.video_resize(*inputs, **kwargs)¶

Loads, decodes and resizes video files with FFmpeg and NVDECODE, which is NVIDIA GPU’s hardware-accelerated video decoding.

The video streams can be in most of the container file formats. FFmpeg is used to parse video containers and returns a batch of sequences with shape (N, F, H, W, C), with N being the batch size, and F the number of frames in the sequence.

This operator combines the features of nvidia.dali.fn.video_reader() and nvidia.dali.fn.resize().

Note

The decoder supports only constant frame-rate videos.

Supported backends

‘gpu’

Keyword Arguments

sequence_length (int) – Frames to load per sequence.
additional_decode_surfaces (int, optional, default = 2) –
Additional decode surfaces to use beyond minimum required.

This argument is ignored when the decoder cannot determine the minimum number of decode surfaces

Note

This can happen when the driver is an older version.

This parameter can be used to trade off memory usage with performance.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
channels (int, optional, default = 3) – Number of channels.
dont_use_mmap (bool, optional, default = False) –
If set to True, the Loader will use plain file I/O instead of trying to map the file in memory.

Mapping provides a small performance benefit when accessing a local file system, but most network file systems, do not provide optimum performance.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.UINT8) –
Output data type.

Supported types: UINT8 or FLOAT.
enable_frame_num (bool, optional, default = False) – If the file_list or filenames argument is passed, returns the frame number output.
enable_timestamps (bool, optional, default = False) – If the file_list or filenames argument is passed, returns the timestamps output.
file_list (str, optional, default = ‘’) –
Path to the file with a list of file label [start_frame [end_frame]] values.

Positive value means the exact frame, negative counts as a Nth frame from the end (it follows python array indexing schema), equal values for the start and end frame would yield an empty sequence and a warning. This option is mutually exclusive with filenames and file_root.
file_list_frame_num (bool, optional, default = False) –
If the start/end timestamps are provided in file_list, you can interpret them as frame numbers instead of as timestamps.

If floating point values have been provided, the start frame number will be rounded up and the end frame number will be rounded down.

Frame numbers start from 0.
file_list_include_preceding_frame (bool, optional, default = False) –
Changes the behavior how file_list start and end frame timestamps are translated to a frame number.

If the start/end timestamps are provided in file_list as timestamps, the start frame is calculated as ceil(start_time_stamp * FPS) and the end as floor(end_time_stamp * FPS). If this argument is set to True, the equation changes to floor(start_time_stamp * FPS) and ceil(end_time_stamp * FPS) respectively. In effect, the first returned frame is not later, and the end frame not earlier, than the provided timestamps. This behavior is more aligned with how the visible timestamps are correlated with displayed video frames.

Note

When file_list_frame_num is set to True, this option does not take any effect.

Warning

This option is available for legacy behavior compatibility.
file_root (str, optional, default = ‘’) –
Path to a directory that contains the data files.

This option is mutually exclusive with filenames and file_list.
filenames (str or list of str, optional, default = []) –
File names of the video files to load.

This option is mutually exclusive with file_list and file_root.
image_type (nvidia.dali.types.DALIImageType, optional, default = DALIImageType.RGB) – The color space of the output frames (RGB or YCbCr).
initial_fill (int, optional, default = 1024) –
Size of the buffer that is used for shuffling.

If random_shuffle is False, this parameter is ignored.
interp_type (nvidia.dali.types.DALIInterpType or TensorList of nvidia.dali.types.DALIInterpType, optional, default = DALIInterpType.INTERP_LINEAR) –
Type of interpolation to be used.

Use min_filter and mag_filter to specify different filtering for downscaling and upscaling.
labels (int or list of int, optional) –
Labels associated with the files listed in filenames argument.

If an empty list is provided, sequential 0-based indices are used as labels. If not provided, no labels will be yielded.
lazy_init (bool, optional, default = False) – Parse and prepare the dataset metadata only during the first run instead of in the constructor.
mag_filter (nvidia.dali.types.DALIInterpType or TensorList of nvidia.dali.types.DALIInterpType, optional, default = DALIInterpType.INTERP_LINEAR) – Filter used when scaling up.
max_size (float or list of float, optional) –
Limit of the output size.

When the operator is configured to keep aspect ratio and only the smaller dimension is specified, the other(s) can grow very large. This can happen when using resize_shorter argument or “not_smaller” mode or when some extents are left unspecified.

This parameter puts a limit to how big the output can become. This value can be specified per-axis or uniformly for all axes.

Note

When used with “not_smaller” mode or resize_shorter argument, max_size takes precedence and the aspect ratio is kept - for example, resizing with mode="not_smaller", size=800, max_size=1400 an image of size 1200x600 would be resized to 1400x700.
min_filter (nvidia.dali.types.DALIInterpType or TensorList of nvidia.dali.types.DALIInterpType, optional, default = DALIInterpType.INTERP_LINEAR) – Filter used when scaling down.
minibatch_size (int, optional, default = 32) – Maximum number of images that are processed in a kernel call.
mode (str, optional, default = ‘default’) –
Resize mode.

Here is a list of supported modes:
- "default" - image is resized to the specified size.
  
  Missing extents are scaled with the average scale of the provided ones.
- "stretch" - image is resized to the specified size.
  
  Missing extents are not scaled at all.
- "not_larger" - image is resized, keeping the aspect ratio, so that no extent of the output image exceeds the specified size.
  
  For example, a 1280x720, with a desired output size of 640x480, actually produces a 640x360 output.
- "not_smaller" - image is resized, keeping the aspect ratio, so that no extent of the output image is smaller than specified.
  
  For example, a 640x480 image with a desired output size of 1920x1080, actually produces a 1920x1440 output.
  
  This argument is mutually exclusive with resize_longer and resize_shorter
normalized (bool, optional, default = False) – Gets the output as normalized data.
num_shards (int, optional, default = 1) –
Partitions the data into the specified number of parts (shards).

This is typically used for multi-GPU or multi-node training.
pad_last_batch (bool, optional, default = False) –
If set to True, pads the shard by repeating the last sample.

Note

If the number of batches differs across shards, this option can cause an entire batch of repeated samples to be added to the dataset.
pad_sequences (bool, optional, default = False) –
Allows creation of incomplete sequences if there is an insufficient number of frames at the very end of the video.

Redundant frames are zeroed. Corresponding time stamps and frame numbers are set to -1.
prefetch_queue_depth (int, optional, default = 1) –
Specifies the number of batches to be prefetched by the internal Loader.

This value should be increased when the pipeline is CPU-stage bound, trading memory consumption for better interleaving with the Loader thread.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
random_shuffle (bool, optional, default = False) –
Determines whether to randomly shuffle data.

A prefetch buffer with a size equal to initial_fill is used to read data sequentially, and then samples are selected randomly to form a batch.
read_ahead (bool, optional, default = False) –
Determines whether the accessed data should be read ahead.

For large files such as LMDB, RecordIO, or TFRecord, this argument slows down the first access but decreases the time of all of the following accesses.
resize_longer (float or TensorList of float, optional, default = 0.0) –
The length of the longer dimension of the resized image.

This option is mutually exclusive with resize_shorter and explicit size arguments, and the operator keeps the aspect ratio of the original image. This option is equivalent to specifying the same size for all dimensions and mode="not_larger".
resize_shorter (float or TensorList of float, optional, default = 0.0) –
The length of the shorter dimension of the resized image.

This option is mutually exclusive with resize_longer and explicit size arguments, and the operator keeps the aspect ratio of the original image. This option is equivalent to specifying the same size for all dimensions and mode="not_smaller". The longer dimension can be bounded by setting the max_size argument. See max_size argument doc for more info.
resize_x (float or TensorList of float, optional, default = 0.0) –
The length of the X dimension of the resized image.

This option is mutually exclusive with resize_shorter, resize_longer and size. If the resize_y is unspecified or 0, the operator keeps the aspect ratio of the original image. A negative value flips the image.
resize_y (float or TensorList of float, optional, default = 0.0) –
The length of the Y dimension of the resized image.

This option is mutually exclusive with resize_shorter, resize_longer and size. If the resize_x is unspecified or 0, the operator keeps the aspect ratio of the original image. A negative value flips the image.
resize_z (float or TensorList of float, optional, default = 0.0) –
The length of the Z dimension of the resized volume.

This option is mutually exclusive with resize_shorter, resize_longer and size. If the resize_x and resize_y are left unspecified or 0, then the op will keep the aspect ratio of the original volume. Negative value flips the volume.
roi_end (float or list of float or TensorList of float, optional) –
End of the input region of interest (ROI).

Must be specified together with roi_start. The coordinates follow the tensor shape order, which is the same as size. The coordinates can be either absolute (in pixels, which is the default) or relative (0..1), depending on the value of relative_roi argument. If the ROI origin is greater than the ROI end in any dimension, the region is flipped in that dimension.
roi_relative (bool, optional, default = False) – If true, ROI coordinates are relative to the input size, where 0 denotes top/left and 1 denotes bottom/right
roi_start (float or list of float or TensorList of float, optional) –
Origin of the input region of interest (ROI).

Must be specified together with roi_end. The coordinates follow the tensor shape order, which is the same as size. The coordinates can be either absolute (in pixels, which is the default) or relative (0..1), depending on the value of relative_roi argument. If the ROI origin is greater than the ROI end in any dimension, the region is flipped in that dimension.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shard_id (int, optional, default = 0) – Index of the shard to read.
size (float or list of float or TensorList of float, optional) –
The desired output size.

Must be a list/tuple with one entry per spatial dimension, excluding video frames and channels. Dimensions with a 0 extent are treated as absent, and the output size will be calculated based on other extents and mode argument.
skip_cached_images (bool, optional, default = False) –
If set to True, the loading data will be skipped when the sample is in the decoder cache.

In this case, the output of the loader will be empty.
skip_vfr_check (bool, optional, default = False) –
Skips the check for the variable frame rate (VFR) videos.

Use this flag to suppress false positive detection of VFR videos.

Warning

When the dataset indeed contains VFR files, setting this flag may cause the decoder to malfunction.
step (int, optional, default = -1) –
Frame interval between each sequence.

When the value is less than 0, step is set to sequence_length.
stick_to_shard (bool, optional, default = False) –
Determines whether the reader should stick to a data shard instead of going through the entire dataset.

If decoder caching is used, it significantly reduces the amount of data to be cached, but might affect accuracy of the training.
stride (int, optional, default = 1) – Distance between consecutive frames in the sequence.
subpixel_scale (bool, optional, default = True) –
If True, fractional sizes, directly specified or calculated, will cause the input ROI to be adjusted to keep the scale factor.

Otherwise, the scale factor will be adjusted so that the source image maps to the rounded output size.
temp_buffer_hint (int, optional, default = 0) –
Initial size in bytes, of a temporary buffer for resampling.

Note

This argument is ignored for the CPU variant.
tensor_init_bytes (int, optional, default = 1048576) – Hint for how much memory to allocate per image.

nvidia.dali.fn.readers.webdataset(*inputs, **kwargs)¶

A reader for the webdataset format.

The webdataset format is a way of providing efficient access to datasets stored in tar archives.

Storing data in POSIX tar archives greatly speeds up I/O operations on mechanical storage devices and on network file systems because it allows the operating system to reduce the number of I/O operations and to read the data ahead.

WebDataset fulfils a similar function to Tensorflow’s TFRecord/tf.Example classes, but is much easier to adopt because it does not actually require any data conversion. The data is stored in exactly the same format inside tar files as it is on disk, and all preprocessing and data augmentation code remains unchanged.

The dataset consists of one or more tar archives, each of which is further split into samples. A sample contains one or more components that correspond to the actual files contained within the archive. The components that belong to a specific sample are aggregated by filename without extension (for the specifics about the extensions please read the description of the ext parameter below). Note that samples with their filename starting with a dot will not be loaded, as well as entries that are not regular files.

In addition to the tar archive with data, each archive should come with a corresponding index file. The index file can be generated using a dedicated script:

<path_to_dali>/tools/wds2idx.py <path_to_archive> <path_to_index_file>

If the index file is not provided, it will be automatically inferred from the tar file. Keep in mind though that it will add considerable startup time for big datasets.

The format of the index file is:

v1.2 <num_samples>
<component1_ext> <component1_data_offset> <component1_size> <component2_ext> <component2_data_offset> <component2_size> ...
...

Based on https://github.com/webdataset/webdataset

Supported backends

‘cpu’

Keyword Arguments

ext (str or list of str) –
The extension sets for each of the outputs produced.

The number of extension sets determines the number of outputs of the reader. The extensions of the components are counted as the text after the first dot in the name of the file (excluding the samples starting with a dot). The different extension options should be separated with a semicolon (‘;’) and may contain dots.

Example: “left.png;right.jpg”
paths (str or list of str) –
The list of (one or more) paths to the webdataset archives.

Has to be the same length as the index_paths argument.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dont_use_mmap (bool, optional, default = False) –
If set to True, the Loader will use plain file I/O instead of trying to map the file in memory.

Mapping provides a small performance benefit when accessing a local file system, but most network file systems, do not provide optimum performance.
dtypes (nvidia.dali.types.DALIDataType or list of nvidia.dali.types.DALIDataType, optional) –
Data types of the respective outputs.

The default output data types are UINT8. However, if set, each output data type should be specified. Moreover, the tar file should be constructed so that it will only output a sample with its byte size divisible by the size of the data type.
index_paths (str or list of str, optional, default = []) –
The list of the index files corresponding to the respective webdataset archives.

Has to be the same length as the paths argument. In case it is not provided, it will be inferred automatically from the webdataset archive.
initial_fill (int, optional, default = 1024) –
Size of the buffer that is used for shuffling.

If random_shuffle is False, this parameter is ignored.
lazy_init (bool, optional, default = False) – Parse and prepare the dataset metadata only during the first run instead of in the constructor.
missing_component_behavior (str, optional, default = ‘’) –
Specifies what to do in case there is not any file in a sample corresponding to a certain output.
Possible behaviors:
- ”empty” (default) - in that case the output that was not set will just contain an empty tensor
- ”skip” - in that case the entire sample will just be skipped (no penalty to performance except for reduced caching of the archive)
- ”error” - in that case an exception will be raised and te execution stops
num_shards (int, optional, default = 1) –
Partitions the data into the specified number of parts (shards).

This is typically used for multi-GPU or multi-node training.
pad_last_batch (bool, optional, default = False) –
If set to True, pads the shard by repeating the last sample.

Note

If the number of batches differs across shards, this option can cause an entire batch of repeated samples to be added to the dataset.
prefetch_queue_depth (int, optional, default = 1) –
Specifies the number of batches to be prefetched by the internal Loader.

This value should be increased when the pipeline is CPU-stage bound, trading memory consumption for better interleaving with the Loader thread.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
random_shuffle (bool, optional, default = False) –
Determines whether to randomly shuffle data.

A prefetch buffer with a size equal to initial_fill is used to read data sequentially, and then samples are selected randomly to form a batch.
read_ahead (bool, optional, default = False) –
Determines whether the accessed data should be read ahead.

For large files such as LMDB, RecordIO, or TFRecord, this argument slows down the first access but decreases the time of all of the following accesses.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shard_id (int, optional, default = 0) – Index of the shard to read.
skip_cached_images (bool, optional, default = False) –
If set to True, the loading data will be skipped when the sample is in the decoder cache.

In this case, the output of the loader will be empty.
stick_to_shard (bool, optional, default = False) –
Determines whether the reader should stick to a data shard instead of going through the entire dataset.

If decoder caching is used, it significantly reduces the amount of data to be cached, but might affect accuracy of the training.
tensor_init_bytes (int, optional, default = 1048576) – Hint for how much memory to allocate per image.

nvidia.dali.fn.reductions¶

nvidia.dali.fn.reductions.max(*inputs, **kwargs)¶

Gets maximal input element along provided axes.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

axes (int or list of int, optional) –
Axis or axes along which reduction is performed.

Not providing any axis results in reduction of all elements.
axis_names (layout str, optional) –
Name(s) of the axis or axes along which the reduction is performed.

The input layout is used to translate the axis names to axis indices, for example axis_names="HW" with input layout “FHWC” is equivalent to specifying axes=[1,2]. This argument cannot be used together with axes.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
keep_dims (bool, optional, default = False) – If True, maintains original input dimensions.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.reductions.mean(*inputs, **kwargs)¶

Gets mean of elements along provided axes.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

axes (int or list of int, optional) –
Axis or axes along which reduction is performed.

Not providing any axis results in reduction of all elements.
axis_names (layout str, optional) –
Name(s) of the axis or axes along which the reduction is performed.

The input layout is used to translate the axis names to axis indices, for example axis_names="HW" with input layout “FHWC” is equivalent to specifying axes=[1,2]. This argument cannot be used together with axes.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.NO_TYPE) – Output data type. This type is used to accumulate the result.
keep_dims (bool, optional, default = False) – If True, maintains original input dimensions.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.reductions.mean_square(*inputs, **kwargs)¶

Gets mean square of elements along provided axes.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

axes (int or list of int, optional) –
Axis or axes along which reduction is performed.

Not providing any axis results in reduction of all elements.
axis_names (layout str, optional) –
Name(s) of the axis or axes along which the reduction is performed.

The input layout is used to translate the axis names to axis indices, for example axis_names="HW" with input layout “FHWC” is equivalent to specifying axes=[1,2]. This argument cannot be used together with axes.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.NO_TYPE) – Output data type. This type is used to accumulate the result.
keep_dims (bool, optional, default = False) – If True, maintains original input dimensions.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.reductions.min(*inputs, **kwargs)¶

Gets minimal input element along provided axes.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

axes (int or list of int, optional) –
Axis or axes along which reduction is performed.

Not providing any axis results in reduction of all elements.
axis_names (layout str, optional) –
Name(s) of the axis or axes along which the reduction is performed.

The input layout is used to translate the axis names to axis indices, for example axis_names="HW" with input layout “FHWC” is equivalent to specifying axes=[1,2]. This argument cannot be used together with axes.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
keep_dims (bool, optional, default = False) – If True, maintains original input dimensions.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.reductions.rms(*inputs, **kwargs)¶

Gets root mean square of elements along provided axes.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

axes (int or list of int, optional) –
Axis or axes along which reduction is performed.

Not providing any axis results in reduction of all elements.
axis_names (layout str, optional) –
Name(s) of the axis or axes along which the reduction is performed.

The input layout is used to translate the axis names to axis indices, for example axis_names="HW" with input layout “FHWC” is equivalent to specifying axes=[1,2]. This argument cannot be used together with axes.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.NO_TYPE) – Output data type. This type is used to accumulate the result.
keep_dims (bool, optional, default = False) – If True, maintains original input dimensions.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.reductions.std_dev(*inputs, **kwargs)¶

Gets standard deviation of elements along provided axes.

Supported backends

‘cpu’
‘gpu’

Parameters

input0 (TensorList) – Input to the operator.
input1 (TensorList) – Input to the operator.

Keyword Arguments

axes (int or list of int, optional) –
Axis or axes along which reduction is performed.

Not providing any axis results in reduction of all elements.
axis_names (layout str, optional) –
Name(s) of the axis or axes along which the reduction is performed.

The input layout is used to translate the axis names to axis indices, for example axis_names="HW" with input layout “FHWC” is equivalent to specifying axes=[1,2]. This argument cannot be used together with axes.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
ddof (int, optional, default = 0) – Delat Degrees of Freedom. Adjusts the divisor used in calculations, which is N - ddof.
keep_dims (bool, optional, default = False) – If True, maintains original input dimensions.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.reductions.sum(*inputs, **kwargs)¶

Gets sum of elements along provided axes.

Supported backends

‘cpu’
‘gpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

axes (int or list of int, optional) –
Axis or axes along which reduction is performed.

Not providing any axis results in reduction of all elements.
axis_names (layout str, optional) –
Name(s) of the axis or axes along which the reduction is performed.

The input layout is used to translate the axis names to axis indices, for example axis_names="HW" with input layout “FHWC” is equivalent to specifying axes=[1,2]. This argument cannot be used together with axes.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
dtype (nvidia.dali.types.DALIDataType, optional, default = DALIDataType.NO_TYPE) – Output data type. This type is used to accumulate the result.
keep_dims (bool, optional, default = False) – If True, maintains original input dimensions.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.reductions.variance(*inputs, **kwargs)¶

Gets variance of elements along provided axes.

Supported backends

‘cpu’
‘gpu’

Parameters

input0 (TensorList) – Input to the operator.
input1 (TensorList) – Input to the operator.

Keyword Arguments

axes (int or list of int, optional) –
Axis or axes along which reduction is performed.

Not providing any axis results in reduction of all elements.
axis_names (layout str, optional) –
Name(s) of the axis or axes along which the reduction is performed.

The input layout is used to translate the axis names to axis indices, for example axis_names="HW" with input layout “FHWC” is equivalent to specifying axes=[1,2]. This argument cannot be used together with axes.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
ddof (int, optional, default = 0) – Delat Degrees of Freedom. Adjusts the divisor used in calculations, which is N - ddof.
keep_dims (bool, optional, default = False) – If True, maintains original input dimensions.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.segmentation¶

nvidia.dali.fn.segmentation.random_mask_pixel(*inputs, **kwargs)¶

Selects random pixel coordinates in a mask, sampled from a uniform distribution.

Based on run-time argument foreground, it returns either only foreground pixels or any pixels.

Pixels are classificed as foreground either when their value exceeds a given threshold or when it’s equal to a specific value.

Supported backends

‘cpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
foreground (int or TensorList of int, optional, default = 0) –
If different than 0, the pixel position is sampled uniformly from all foreground pixels.

If 0, the pixel position is sampled uniformly from all available pixels.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
threshold (float or TensorList of float, optional, default = 0.0) –
All pixels with a value above this threshold are interpreted as foreground.

This argument is mutually exclusive with value argument.
value (int or TensorList of int, optional) –
All pixels equal to this value are interpreted as foreground.

This argument is mutually exclusive with threshold argument and is meant to be used only with integer inputs.

nvidia.dali.fn.segmentation.random_object_bbox(*inputs, **kwargs)¶

Randomly selects an object from a mask and returns its bounding box.

This operator takes a labeled segmentation map as its input. With probability foreground_prob it randomly selects a label (uniformly or according to the distribution given as class_weights), extracts connected blobs of pixels with the selected label and randomly selects one of the blobs. The blobs may be further filtered according to k_largest and threshold. The output is a bounding box of the selected blob in one of the formats described in format.

With probability 1-foreground_prob, the entire area of the input is returned.

Supported backends

‘cpu’

Parameters

input (TensorList) – Input to the operator.

Keyword Arguments

background (int or TensorList of int, optional, default = 0) –
Background label.

If left unspecified, it’s either 0 or any value not in classes.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
cache_objects (bool, optional, default = False) –
Cache object bounding boxes to avoid the computational cost of finding object blobs in previously seen inputs.

Searching for blobs of connected pixels and finding boxes can take a long time. When the dataset has few items, but item size is big, you can use caching to save the boxes and reuse them when the same input is seen again. The inputs are compared based on 256-bit hash, which is much faster to compute than to recalculate the object boxes.
class_weights (float or list of float or TensorList of float, optional) –
Relative probabilities of foreground classes.

Each value corresponds to a class label in classes. If classes are not specified, consecutive 1-based labels are assigned.

The sum of the weights doesn’t have to be equal to 1 - if it isn’t the weights will be normalized .
classes (int or list of int or TensorList of int, optional) –
List of labels considered as foreground.

If left unspecified, all labels not equal to background are considered foreground.
foreground_prob (float or TensorList of float, optional, default = 1.0) – Probability of selecting a foreground bounding box.
format (str, optional, default = ‘anchor_shape’) –
Format in which the data is returned.
Possible choices are::
- ”anchor_shape” (the default) - there are two outputs: anchor and shape
- ”start_end” - there are two outputs: bounding box start and one-past-end coordinates
- ”box” - there is one output that contains concatenated start and end coordinates
ignore_class (bool, optional, default = False) –
If True, all objects are picked with equal probability, regardless of the class they belong to. Otherwise, a class is picked first and then an object is randomly selected from this class.

This argument is incompatible with classes, class_weights or output_class.

Note

This flag only affects the probability with which blobs are selected. It does not cause blobs of different classes to be merged.
k_largest (int, optional) –
If specified, the boxes are sorted by decreasing volume and only k_largest are considered.

If ignore_class is True, k_largest referes to all boxes; otherwise it refers to the selected class.
output_class (bool, optional, default = False) –
If True, an additional output is produced which contains the label of the class to which the selected box belongs, or background label if the selected box is not an object bounding box.
The output may not be an object bounding box when any of the following conditions occur:
- the sample was randomly (according to foreground_prob) chosen not be be a foreground one
- the sample contained no foreground objects
- no bounding box met the required size threshold.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
threshold (int or list of int or TensorList of int, optional) –
Per-axis minimum size of the bounding boxes to return.

If the selected class doesn’t contain any bounding box that meets this condition, it is rejected and another class is picked. If no class contains a satisfactory box, the entire input area is returned.

nvidia.dali.fn.segmentation.select_masks(*inputs, **kwargs)¶

Selects a subset of polygons by their mask ids.

The operator expects three inputs describing multiple segmentation mask polygons belonging to different mask ids and a list of selected mask ids.

Each sample can contain several polygons belonging to different masks, and each polygon can be composed by an arbitrary number of vertices (at least 3). The masks polygons are described by the inputs polygons and vertices and the operator produces output polygons and vertices where only the polygons associated with the selected masks are present.

Note

The format of polygons and vertices is the same as produced by COCOReader.

Examples:

Let us assume the following input mask, where symbolic coordinates are used for a clearer example:

polygons = [[0, 0, 3], [1, 3, 7], [2, 7, 10]]
vertices = [[x0, y0], [x1, y1], [x2, y2], [x3, y3], [x4, y4], [x5, y5], [x6, y6], [x7, y7], [x8, y8], [x9, y9]]

Example 1: Selecting a single mask with id 1, maintaining the original id:

mask_ids = [1], ``reindex_masks`` = False
out_polygons = [[1, 0, 4]]
out_vertices = [[x3, y3], [x4, y4], [x5, y5], [x6, y6]]

Example 2: Selecting two out of the three masks, replacing the mask ids with the indices at which they appeared in mask_ids input:

mask_ids = [2, 0]
reindex_masks = True
out_polygons = [[0, 3, 6], [1, 0, 3]]
out_vertices = [[x0, y0], [x1, y1], [x2, y2], [x7, y7], [x8, y8], [x9, y9]]

Supported backends

‘cpu’

Parameters

mask_ids (1D TensorList of int) – List of identifiers of the masks to be selected. The list should not contain duplicates.
polygons (2D TensorList of int) –
Polygons, described by 3 columns:
```
[[mask_id0, start_vertex_idx0, end_vertex_idx0],
 [mask_id1, start_vertex_idx1, end_vertex_idx1],
 ...,
 [mask_idn, start_vertex_idxn, end_vertex_idxn],]
```
with mask_id being the identifier of the mask this polygon belongs to, and [start_vertex_idx, end_vertex_idx) describing the range of indices from vertices that belong to this polygon.
vertices (2D TensorList) –
Vertex data stored in interleaved format:
```
[[x0, y0, ...],
 [x1, y1, ...],
 ... ,
 [xn, yn, ...]]
```
The operator accepts vertices with arbitrary number of coordinates.

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
reindex_masks (bool, optional, default = False) – If set to True, the output mask ids are replaced with the indices at which they appeared in mask_ids input.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.transforms¶

All operators in this module support only CPU device as they are meant to be provided as an input to named keyword operator arguments. Check for more details the relevant pipeline documentation section.

nvidia.dali.fn.transforms.combine(*inputs, **kwargs)¶

Combines two or more affine transforms.

By default, the transforms are combined such that applying the resulting transform to a point is equivalent to

applying the input transforms in the order as listed.

Example: combining [T1, T2, T3] is equivalent to T3(T2(T1(…))) for default order and equivalent to T1(T2(T3(…)))

for reversed order.

Supported backends

‘cpu’

Parameters

input0 (TensorList) – Input to the operator.
input1 (TensorList) – Input to the operator.
input[2..98] (TensorList, optional) – This function accepts up to 97 optional positional inputs

Keyword Arguments

bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
reverse_order (bool, optional, default = False) –
Determines the order when combining affine transforms.

If set to False (default), the operator’s affine transform will be applied to the input transform. If set to True, the input transform will be applied to the operator’s transform.

If there’s no input, this argument is ignored.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.transforms.crop(*inputs, **kwargs)¶

Produces an affine transform matrix that maps a reference coordinate space to another one.

This transform can be used to adjust coordinates after a crop operation so that a from_start point will be mapped to to_start and from_end will be mapped to to_end.

If another transform matrix is passed as an input, the operator applies the transformation to the matrix provided.

Note

The output of this operator can be fed directly to CoordTransform and WarpAffine operators.

Supported backends

‘cpu’

Parameters

input (TensorList, optional) – Input to the operator.

Keyword Arguments

absolute (bool, optional, default = False) – If set to true, start and end coordinates will be swapped if start > end.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
from_end (float or list of float or TensorList of float, optional, default = [1.0]) –
The upper bound of the original coordinate space.

Note

If left empty, a vector of ones will be assumed. If a single value is provided, it will be repeated to match the number of dimensions
from_start (float or list of float or TensorList of float, optional, default = [0.0]) –
The lower bound of the original coordinate space.

Note

If left empty, a vector of zeros will be assumed. If a single value is provided, it will be repeated to match the number of dimensions
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
reverse_order (bool, optional, default = False) –
Determines the order when combining affine transforms.

If set to False (default), the operator’s affine transform will be applied to the input transform. If set to True, the input transform will be applied to the operator’s transform.

If there’s no input, this argument is ignored.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
to_end (float or list of float or TensorList of float, optional, default = [1.0]) –
The upper bound of the destination coordinate space.

Note

If left empty, a vector of ones will be assumed. If a single value is provided, it will be repeated to match the number of dimensions
to_start (float or list of float or TensorList of float, optional, default = [0.0]) –
The lower bound of the destination coordinate space.

Note

If left empty, a vector of zeros will be assumed. If a single value is provided, it will be repeated to match the number of dimensions

nvidia.dali.fn.transforms.rotation(*inputs, **kwargs)¶

Produces a rotation affine transform matrix.

If another transform matrix is passed as an input, the operator applies rotation to the matrix provided.

The number of dimensions is assumed to be 3 if a rotation axis is provided or 2 otherwise.

Note

The output of this operator can be fed directly to CoordTransform and WarpAffine operators.

Supported backends

‘cpu’

Parameters

input (TensorList, optional) – Input to the operator.

Keyword Arguments

angle (float or TensorList of float) – Angle, in degrees.
axis (float or list of float or TensorList of float, optional) –
Axis of rotation (applies only to 3D transforms).

The vector does not need to be normalized, but it must have a non-zero length.

Reversing the vector is equivalent to changing the sign of angle.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
center (float or list of float or TensorList of float, optional) –
The center of the rotation.

If provided, the number of elements should match the dimensionality of the transform.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
reverse_order (bool, optional, default = False) –
Determines the order when combining affine transforms.

If set to False (default), the operator’s affine transform will be applied to the input transform. If set to True, the input transform will be applied to the operator’s transform.

If there’s no input, this argument is ignored.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.transforms.scale(*inputs, **kwargs)¶

Produces a scale affine transform matrix.

If another transform matrix is passed as an input, the operator applies scaling to the matrix provided.

Note

The output of this operator can be fed directly to CoordTransform and WarpAffine operators.

Supported backends

‘cpu’

Parameters

input (TensorList, optional) – Input to the operator.

Keyword Arguments

scale (float or list of float or TensorList of float) –
The scale factor, per dimension.

The number of dimensions of the transform is inferred from this argument.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
center (float or list of float or TensorList of float, optional) –
The center of the scale operation.

If provided, the number of elements should match the one of scale argument.
ndim (int, optional) –
Number of dimensions.

It should be provided when the number of dimensions can’t be inferred. For example, when scale is a scalar value and there’s no input transform.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
reverse_order (bool, optional, default = False) –
Determines the order when combining affine transforms.

If set to False (default), the operator’s affine transform will be applied to the input transform. If set to True, the input transform will be applied to the operator’s transform.

If there’s no input, this argument is ignored.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.fn.transforms.shear(*inputs, **kwargs)¶

Produces a shear affine transform matrix.

If another transform matrix is passed as an input, the operator applies the shear mapping to the matrix provided.

Note

The output of this operator can be fed directly to CoordTransform and WarpAffine operators.

Supported backends

‘cpu’

Parameters

input (TensorList, optional) – Input to the operator.

Keyword Arguments

angles (float or list of float or TensorList of float, optional) –
The shear angles, in degrees.

This argument is mutually exclusive with shear.

For 2D, angles contains two elements: angle_x, angle_y.

For 3D, angles contains six elements: angle_xy, angle_xz, angle_yx, angle_yz, angle_zx, angle_zy.

A shear angle is translated to a shear factor as follows:
```
shear_factor = tan(deg2rad(shear_angle))
```
Note

The valid range of values is between -90 and 90 degrees. This argument is mutually exclusive with shear. If provided, the number of dimensions of the transform is inferred from this argument.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
center (float or list of float or TensorList of float, optional) –
The center of the shear operation.

If provided, the number of elements should match the dimensionality of the transform.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
reverse_order (bool, optional, default = False) –
Determines the order when combining affine transforms.

If set to False (default), the operator’s affine transform will be applied to the input transform. If set to True, the input transform will be applied to the operator’s transform.

If there’s no input, this argument is ignored.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
shear (float or list of float or TensorList of float, optional) –
The shear factors.

For 2D, shear contains two elements: shear_x, shear_y.

For 3D, shear contains six elements: shear_xy, shear_xz, shear_yx, shear_yz, shear_zx, shear_zy.

A shear factor value can be interpreted as the offset to be applied in the first axis when moving in the direction of the second axis.

Note

This argument is mutually exclusive with angles. If provided, the number of dimensions of the transform is inferred from this argument.

nvidia.dali.fn.transforms.translation(*inputs, **kwargs)¶

Produces a translation affine transform matrix.

If another transform matrix is passed as an input, the operator applies translation to the matrix provided.

Note

The output of this operator can be fed directly to CoordTransform and WarpAffine operators.

Supported backends

‘cpu’

Parameters

input (TensorList, optional) – Input to the operator.

Keyword Arguments

offset (float or list of float or TensorList of float) –
The translation vector.

The number of dimensions of the transform is inferred from this argument.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
reverse_order (bool, optional, default = False) –
Determines the order when combining affine transforms.

If set to False (default), the operator’s affine transform will be applied to the input transform. If set to True, the input transform will be applied to the operator’s transform.

If there’s no input, this argument is ignored.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.

nvidia.dali.plugin.numba.fn.experimental¶

nvidia.dali.plugin.numba.fn.experimental.numba_function(*inputs, **kwargs)¶

Invokes a njit compiled Numba function.

The run function should be a Python function that can be compiled in Numba nopython mode. A function taking a single input and producing a single output should follow the following definition:

def run_fn(out0, in0)

where out0 and in0 are numpy array views of the input and output tensors. If the operator is configured to run in batch mode, then the first dimension of the arrays is the sample index.

Note that the function can take at most 6 inputs and 6 outputs.

Additionally, an optional setup function calculating the shape of the output so DALI can allocate memory for the output with the following definition:

def setup_fn(outs, ins)

The setup function is invoked once for the whole batch. The first dimension of outs, ins is the number of outputs/inputs, respectively. The second dimension is the sample index. For example, the first sample on the second output can be accessed by outs[1][0].

If no setup function provided, the output shape and data type will be the same as the input.

Note

This operator is experimental and its API might change without notice.

Example 1:

The following example shows a simple setup function which permutes the order of dimensions in the shape.

def setup_change_out_shape(outs, ins):
    out0 = outs[0]
    in0 = ins[0]
    perm = [1, 0, 2]
    for sample_idx in range(len(out0)):
        for d in range(len(perm)):
            out0[sample_idx][d] = in0[sample_idx][perm[d]]

Since the setup function is running for the whole batch, we need to iterate and permute each sample’s shape individually. For shapes = [(10, 20, 30), (20, 10, 30)] it will produce output with shapes = [(20, 10, 30), (10, 20, 30)].

Also lets provide run function:

def run_fn(out0, in0):
    for i in range(in0.shape[0]):
        for j in range(in0.shape[1]):
            out0[j, i] = in0[i, j]

The run function can work per-sample or per-batch, depending on the batch_processing argument.

A run function working per-batch may look like this:

def run_fn(out0_samples, in0_samples):
    for out0, in0 in zip(out0_samples, in0_samples):
        for i in range(in0.shape[0]):
            for j in range(in0.shape[1]):
                out0[j, i] = in0[i, j]

A run function working per-sample may look like this:

def run_fn(out0, in0):
    for i in range(in0.shape[0]):
        for j in range(in0.shape[1]):
            out0[j, i] = in0[i, j]

This operator allows sequence inputs and supports volumetric data.

This operator will not be optimized out of the graph.

Supported backends

‘cpu’

Parameters

input0 (TensorList) – Input to the operator.
input[1..5] (TensorList, optional) – This function accepts up to 5 optional positional inputs

Keyword Arguments

in_types (nvidia.dali.types.DALIDataType or list of nvidia.dali.types.DALIDataType) – Types of inputs.
ins_ndim (int or list of int) – Number of dimensions which inputs shapes should have.
out_types (nvidia.dali.types.DALIDataType or list of nvidia.dali.types.DALIDataType) – Types of outputs.
outs_ndim (int or list of int) – Number of dimensions which outputs shapes should have.
run_fn (object) – Function to be invoked. This function must work in Numba nopython mode.
batch_processing (bool, optional, default = False) –
Determines whether the function is invoked once per batch or separately for each sample in the batch.

When batch_processing is set to True, the function processes the whole batch. It is necessary if the function has to perform cross-sample operations and may be beneficial if significant part of the work can be reused. For other use cases, specifying False and using per-sample processing function allows the operator to process samples in parallel.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.
setup_fn (object, optional) – Setup function setting shapes for outputs. This function is invoked once per batch. Also this function must work in Numba nopython mode.

nvidia.dali.plugin.pytorch.fn¶

nvidia.dali.plugin.pytorch.fn.torch_python_function(*inputs, **kwargs)¶

Executes a function that is operating on Torch tensors.

This class is analogous to nvidia.dali.fn.python_function() but the tensor data is handled as PyTorch tensors.

This operator allows sequence inputs and supports volumetric data.

This operator will not be optimized out of the graph.

Supported backends

‘cpu’
‘gpu’

Parameters

input[0..255] (TensorList, optional) – This function accepts up to 256 optional positional inputs

Keyword Arguments

function (object) – Function object.
batch_processing (bool, optional, default = False) – Determines whether the function gets an entire batch as an input.
bytes_per_sample_hint (int or list of int, optional, default = [0]) –
Output size hint, in bytes per sample.

If specified, the operator’s outputs residing in GPU or page-locked host memory will be preallocated to accommodate a batch of samples of this size.
num_outputs (int, optional, default = 1) – Number of outputs.
output_layouts (layout str or list of layout str, optional) –
Tensor data layouts for the outputs.

This argument can be a list that contains a distinct layout for each output. If the list has fewer than num_outputs elements, only the first outputs have the layout set and the rest of the outputs have no layout assigned.
preserve (bool, optional, default = False) – Prevents the operator from being removed from the graph even if its outputs are not used.
seed (int, optional, default = -1) –
Random seed.

If not provided, it will be populated based on the global seed of the pipeline.