nvidia.dali.fn.reinterpret¶
- 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) –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.
The output shape is calculated by multiplying the input shape by
rel_shape
:out_shape[i] = in_shape[i] * rel_shape[i]
An additional argument
src_dims
may be used to alter which source dimension is used for calculating the output shape:out_shape[i] = in_shape[src_dims[i]] * rel_shape[i]
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 arel_shape = [0.5, -1]
results in the output shape[240, 3840]
.The number of dimensions is subject to the following restrictions:
if
src_dims
argument is used, the number of elements insrc_dims
andrel_shape
must matchotherwise, the length of
rel_shape
must not exceed the number of dimensions in the input except when the last element inrel_shape
is negative, in which case an extra dimension at the end will be added
Note
rel_shape
andshape
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.
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]
andshape = [240, -1]
results in the output shape[240, 3840]
.Note
rel_shape
andshape
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 asrc_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, specifyingrel_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.