Slice

Extracts elements from an input tensor into an output tensor.

Attributes

start First coordinate to use.

size Output dimensions.

stride Stride between coordinates.

slice_mode Describes how out-of-bounds coordinates are handled.

Let \(y\) denote the components of an output coordinate and \(d\) denote dimensions of the input tensor. Define \(x_i=y_i \cdot \text{stride}_i + \text{start}_i\).

  • STRICT_BOUNDS \(output[y_i]=input[x_i]\) if \(0 \leq x_i < d_i\). Otherwise issue error.
  • WRAP \(output[y_i]=input[mod(x_i, d_i)]\)
  • CLAMP \(output[y_i]=input[z_i] \text{ where } z_i=\begin{cases}0, \text{if } x_i<0 \\ d_i-1, \text{if } x_i \geq d_i \\ x_i, \text{otherwise}\end{cases}\)
  • FILL \(output[y_i]=\begin{cases}input[x_i], \text{if } 0\leq x_i < d_i \\ \text{fill value}, \text{otherwise}\end{cases}\)
  • REFLECT \(output[y_i]=input[z_i] \text{ where } z_i=\begin{cases}2 \cdot d_i-2-c_i, \text{if } c_i \geq d_i \\ c_i, \text{otherwise } \end{cases} \text{ where } c_i=mod(|x_i|, 2 \cdot d_i -2)\)

Alternatively, the start, size, and stride can be specified by inputs to the slice operation.

Inputs

input0: tensor of type T.

input1: optional tensor of type Int32 or Int64` with ``start.

input2: optional tensor of type Int32 or Int64` with ``size.

input3: optional tensor of type Int32 or Int64` with ``stride.

input4: optional tensor of type T (or type immplicitly convertible to T) containing fill value for FILL mode.

Outputs

output: tensor of type T.

Data Types

T: bool, int8, int32, int64, float16, float32, bfloat16

  • Note: CLAMP, FILL and REFLECT modes do not support int64 or bfloat16.

Shape Information

input0 has shape \([d_0,...,d_{n-1}]\).

input1 has shape \([n]\).

input2 has shape \([n]\).

input3 has shape \([n]\).

input4 has shape \([]\).

output is a tensor with shape \([\text{size}_0,...,\text{size}_{n-1}]\).

DLA Support

DLA FP16 is supported.

Examples

Slice
in1 = network.add_input("input1", dtype=trt.float32, shape=(3, 3))
layer = network.add_slice(in1, start=(0, 0), shape=(2, 2), stride=(1, 1))
network.mark_output(layer.get_output(0))

inputs[in1.name] = np.array([[0, 1, 2], [3, 4, 5], [6, 7, 8]])

outputs[layer.get_output(0).name] = layer.get_output(0).shape
expected[layer.get_output(0).name] = np.array([[0, 1], [3, 4]])
Slice Fill
in1 = network.add_input("input1", dtype=trt.float32, shape=(2, 2))
layer = network.add_slice(in1, start=(0, 0), shape=(3, 3), stride=(1, 1))
fill_constant = network.add_input("fill_constant", dtype=trt.float32, shape=())
layer.mode = trt.SampleMode.FILL
layer.set_input(4, fill_constant)
network.mark_output(layer.get_output(0))

inputs[in1.name] = np.zeros(shape=(2, 2))
inputs[fill_constant.name] = np.array([1.0])

outputs[layer.get_output(0).name] = layer.get_output(0).shape
expected[layer.get_output(0).name] = np.array([[0, 0, 1], [0, 0, 1], [1, 1, 1]])

C++ API

For more information about the C++ ISliceLayer operator, refer to the C++ ISliceLayer documentation.

Python API

For more information about the Python ISliceLayer operator, refer to the Python ISliceLayer documentation.