How to perform element-wise subtraction on tensors in PyTorch?

To perform element-wise subtraction on tensors, we can use the torch.sub() method of PyTorch. The corresponding elements of the tensors are subtracted. We can subtract a scalar or tensor from another tensor with same or different dimensions. The dimension of the final tensor will be the same as the dimension of the higher-dimensional tensor due to PyTorch's broadcasting rules.

Syntax

torch.sub(input, other, *, alpha=1, out=None)

Parameters:

• input ? The tensor to be subtracted from
• other ? The tensor or scalar to subtract
• alpha ? The multiplier for other (default: 1)
• out ? The output tensor (optional)

Subtracting a Scalar from a Tensor

Here are three different ways to subtract a scalar from a tensor ?

import torch

# Create a tensor
t = torch.Tensor([1.5, 2.03, 3.8, 2.9])
print("Original Tensor t:\n", t)

# Method 1: Subtract scalar directly
v = torch.sub(t, 5.60)
print("Element-wise subtraction result:\n", v)

# Method 2: Using a 1-element tensor
t1 = torch.Tensor([5.60])
w = torch.sub(t, t1)
print("Element-wise subtraction result:\n", w)

# Method 3: Using tensor with same shape
t2 = torch.Tensor([5.60, 5.60, 5.60, 5.60])
x = torch.sub(t, t2)
print("Element-wise subtraction result:\n", x)

Original Tensor t:
 tensor([1.5000, 2.0300, 3.8000, 2.9000])
Element-wise subtraction result:
 tensor([-4.1000, -3.5700, -1.8000, -2.7000])
Element-wise subtraction result:
 tensor([-4.1000, -3.5700, -1.8000, -2.7000])
Element-wise subtraction result:
 tensor([-4.1000, -3.5700, -1.8000, -2.7000])

Subtracting 1D Tensor from 2D Tensor

The 1D tensor is broadcast across each row of the 2D tensor ?

import torch

# Create a 2D tensor
T1 = torch.Tensor([[8, 7], [4, 5]])

# Create a 1-D tensor
T2 = torch.Tensor([10, 5])
print("T1:\n", T1)
print("T2:\n", T2)

# Subtract 1-D tensor from 2-D tensor
v = torch.sub(T1, T2)
print("Element-wise subtraction result:\n", v)

T1:
tensor([[8., 7.],
        [4., 5.]])
T2:
 tensor([10., 5.])
Element-wise subtraction result:
tensor([[-2., 2.],
        [-6., 0.]])

Subtracting 2D Tensor from 1D Tensor

The 1D tensor is broadcast to match the 2D tensor dimensions ?

import torch

# Create tensors
T1 = torch.Tensor([[1, 2], [4, 5]])
T2 = torch.Tensor([10, 5])
print("T1:\n", T1)
print("T2:\n", T2)

# Subtract 2-D tensor from 1-D tensor
v = torch.sub(T2, T1)
print("Element-wise subtraction result:\n", v)

T1:
tensor([[1., 2.],
        [4., 5.]])
T2:
 tensor([10., 5.])
Element-wise subtraction result:
tensor([[9., 3.],
        [6., 0.]])

Subtracting 2D Tensors

Element-wise subtraction between tensors of the same dimensions ?

import torch

# Create two 2-D tensors
T1 = torch.Tensor([[8, 7], [3, 4]])
T2 = torch.Tensor([[0, 3], [4, 9]])
print("T1:\n", T1)
print("T2:\n", T2)

# Subtract above two 2-D tensors
v = torch.sub(T1, T2)
print("Element-wise subtraction result:\n", v)

T1:
tensor([[8., 7.],
        [3., 4.]])
T2:
tensor([[0., 3.],
        [4., 9.]])
Element-wise subtraction result:
tensor([[ 8., 4.],
        [-1., -5.]])

Using Alternative Syntax

PyTorch also supports the minus operator (-) for tensor subtraction ?

import torch

a = torch.tensor([5.0, 3.0, 2.0])
b = torch.tensor([1.0, 2.0, 4.0])

# Using torch.sub()
result1 = torch.sub(a, b)
print("Using torch.sub():\n", result1)

# Using minus operator
result2 = a - b
print("Using minus operator:\n", result2)

Using torch.sub():
 tensor([4., 1., -2.])
Using minus operator:
 tensor([4., 1., -2.])

Conclusion

PyTorch provides flexible element-wise subtraction through torch.sub() or the minus operator. Broadcasting allows subtraction between tensors of different dimensions, making operations intuitive and efficient.