How to narrow down a tensor in PyTorch?

torch.narrow() method is used to perform narrow operation on a PyTorch tensor. It returns a new tensor that is a narrowed version of the original input tensor.

For example, a tensor of [4, 3] can be narrowed to a tensor of size [2, 3] or [4, 2]. We can narrow down a tensor along a single dimension at a time. Here, we cannot narrow down both dimensions to a size of [2, 2]. We can also use Tensor.narrow() to narrow down a tensor.

Syntax

torch.narrow(input, dim, start, length)
Tensor.narrow(dim, start, length)

Parameters

• input – It's the PyTorch tensor to narrow.

• dim – It's the dimension along which we have to narrow down the original tensor, input.

• Start – Starting dimension.

• Length – Length to the end dimension from starting dimension.

Steps

• Import the torch library. Make sure you have already installed it.

import torch

• Create a PyTorch tensor and print the tensor and its size.

t = torch.tensor([[1, 2, 3], [4, 5, 6], [7, 8, 9]])
print("Tensor:
", t)
print("Size of tensor:", t.size()) # size 3x3

• Compute torch.narrow(input, dim, start, length) and assign the value to a variable.

t1 = torch.narrow(t, 0, 1, 2)

• Print the resultant tensor and its size, after narrowing.

print("Tensor after Narrowing:
", t2)
print("Size after Narrowing:", t2.size())

Example 1

In the following Python code, the input tensor size is [3, 3]. We use dim = 0, start = 1 and length = 2 to narrow down the tensor along the dimension 0. It returns a new tensor with the dimension [2, 3].

Notice the new tensor is narrowed along the dimension 0 and the length along the dimension 0 is changed to 2.

# import the library
import torch

# create a tensor
t = torch.tensor([[1, 2, 3], [4, 5, 6], [7, 8, 9]])

# print the created tensor
print("Tensor:
", t)
print("Size of Tensor:", t.size())

# Narrow-down the tensor in dimension 0
t1 = torch.narrow(t, 0, 1, 2)
print("Tensor after Narrowing:
", t1)
print("Size after Narrowing:", t1.size())

# Narrow down the tensor in dimension 1
t2 = torch.narrow(t, 1, 1, 2)
print("Tensor after Narrowing:
", t2)
print("Size after Narrowing:", t2.size())

Output

Tensor:
 tensor([[1, 2, 3],
    [4, 5, 6],
    [7, 8, 9]])
Size of Tensor: torch.Size([3, 3])
Tensor after Narrowing:
 tensor([[4, 5, 6],
    [7, 8, 9]])
Size after Narrowing: torch.Size([2, 3])
Tensor after Narrowing:
 tensor([[2, 3],
    [5, 6],
    [8, 9]])
Size after Narrowing: torch.Size([3, 2])

Example 2

The following program shows how to implement the narrow operation using Tensor.narrow().

# import required library
import torch

# create a tensor
t = torch.tensor([[1, 2, 3], [4, 5, 6], [7, 8, 9], [10, 11, 12]])
# print the above created tensor
print("Tensor:
", t)
print("Size of Tensor:", t.size())

# Narrow-down the tensor in dimension 0
t1 = t.narrow(0, 1, 2)
print("Tensor after Narrowing:
", t1)
print("Size after Narrowing:", t1.size())

# Narrow down the tensor in dimension 1
t2 = t.narrow(1, 0, 2)
print("Tensor after Narrowing:
", t2)
print("Size after Narrowing:", t2.size())

Output

Tensor:
 tensor([[ 1, 2, 3],
    [ 4, 5, 6],
    [ 7, 8, 9],
    [10, 11, 12]])
Size of Tensor: torch.Size([4, 3])
Tensor after Narrowing:
 tensor([[4, 5, 6],
    [7, 8, 9]])
Size after Narrowing: torch.Size([2, 3])
Tensor after Narrowing:
 tensor([[ 1, 2],
    [ 4, 5],
    [ 7, 8],
    [10, 11]])
Size after Narrowing: torch.Size([4, 2])