nvidia.dali.fn.erase¶
- 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 ashape
(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
andshape
represent multiple points consecutively (for example,anchor
= (y0, x0, y1, x1, …) andshape
= (h0, w0, h1, w1, …)). Theanchor
andshape
arguments are interpreted based on the value of theaxis_names
argument, or, alternatively, the value of theaxes
argument. If noaxis_names
oraxes
arguments are provided, all dimensions exceptC
(channels) must be specified.Example 1:
anchor
= (10, 20),shape
= (190, 200),axis_names
= “HW”,fill_value
= 0input:
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
= Trueinput:
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
= Falseinput:
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 or TensorList 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.