How to Pad the Input Tensor Boundaries With a Constant Value in PyTorch?

In this article, we will discuss How to Pad the Input Tensor Boundaries With a Constant Value in PyTorch. We can pad the input tensor boundaries with a constant value by using torch.nn.ConstantPad2D() method.

torch.nn.ConstantPad2D() method 

We can pad the boundaries of 3D and 4D tensors and the shape of the input tensor is [C, H, W] and [N, C, H, W] respectively, where N represents the mini-batch size, C represents the number of channels, and  H, W represents the height and width respectively. This method accepts the size of padding and a constant value as input. The boundaries may be the same or different from all sides (left, right, top, bottom). we can increase the height and width of a padded tensor by using top+bottom and left+right respectively. The below syntax is used to pad the input tensor boundaries with a Constant Value.

Syntax: torch.nn.ConstantPad2d(pad, value)

Parameter:

• pad (int, tuple): This is size of padding. The size of padding is an integer or a tuple.
• value: This is constant value. 

Return: This method returns a new tensor with boundaries.

Example 1:

In this example, we will see how to add the same padding sizes to all sides.

Python3

# Import required library
import torch
import torch.nn as nn

# define a tensor
tens = torch.tensor([[[21, 22], [23, 24]]])
print("\n Input Tensor: \n", tens)

# give padding size same for all sides
pad = nn.ConstantPad2d(2, 9)
output = pad(tens)

# display result
print("\n After Pad Input Tensor: \n", output)

Output:

Example 2:

In this example, we will see how to add unique padding sizes to all sides (left, right, top, bottom).

Python3

# Import required library
import torch
import torch.nn as nn

# define a tensor
tens = torch.tensor([[[11, 12], [13, 14]]])
print("\n Input Tensor: \n", tens)

# add unique padding sizes to all sides
# (left, right, top, bottom)
pad = nn.ConstantPad2d((1, 2, 3, 4), 8)
output = pad(tens)

# display result
print("\n After Pad Input Tensor:\n", output)

Output:

Example 3:

In this example, we will see how to pad the boundaries of a batch of tensors.

Python3

# Import required library
import torch
import torch.nn as nn

# define a batch of tensor
tens = torch.tensor([[[11, 12], [13, 14]],
                     [[21, 22], [23, 24]]])

print("\n Input Tensor: \n", tens)

# add unique padding sizes to all sides
# (left, right, top, bottom)
pad = nn.ConstantPad2d(1, 8)
output = pad(tens)

# display result
print("\n After Pad Input Tensor:\n", output)

Output: