How to Create a Zeros Tensor in PyTorch: Simple Guide

In PyTorch, you can create a tensor filled with zeros using torch.zeros(). You specify the shape as a tuple, like torch.zeros((2, 3)), which creates a 2 by 3 tensor filled with zeros.

Syntax

The basic syntax to create a zeros tensor in PyTorch is torch.zeros(size, dtype=None, device=None, requires_grad=False).

• size: A tuple defining the shape of the tensor (e.g., (2, 3) for 2 rows and 3 columns).
• dtype: (Optional) The data type of the tensor elements, like torch.float32 or torch.int64. Defaults to torch.float32.
• device: (Optional) The device where the tensor will be stored, such as CPU or GPU.
• requires_grad: (Optional) If set to True, PyTorch will track operations on the tensor for automatic differentiation.

python

torch.zeros(size, dtype=None, device=None, requires_grad=False)

Example

This example creates a 2x3 tensor filled with zeros and prints it. It also shows how to specify the data type and device.

python

import torch

# Create a 2x3 zeros tensor
zeros_tensor = torch.zeros((2, 3))
print('Zeros tensor on CPU:')
print(zeros_tensor)

# Create a 2x3 zeros tensor with integer type
zeros_int_tensor = torch.zeros((2, 3), dtype=torch.int64)
print('\nZeros tensor with int64 type:')
print(zeros_int_tensor)

# Create a zeros tensor on GPU if available
if torch.cuda.is_available():
    zeros_cuda = torch.zeros((2, 3), device='cuda')
    print('\nZeros tensor on GPU:')
    print(zeros_cuda)
else:
    print('\nGPU not available, skipping GPU tensor creation.')

Output

Zeros tensor on CPU:
tensor([[0., 0., 0.],
        [0., 0., 0.]])

Zeros tensor with int64 type:
tensor([[0, 0, 0],
        [0, 0, 0]])

GPU not available, skipping GPU tensor creation.

Common Pitfalls

Some common mistakes when creating zeros tensors in PyTorch include:

• Forgetting to pass the size as a tuple, e.g., writing torch.zeros(2, 3) instead of torch.zeros((2, 3)). The former works but is less clear and can cause confusion.
• Not specifying dtype when integer zeros are needed, which defaults to floating point.
• Trying to create a tensor on GPU without checking if CUDA is available, which causes errors.

python

import torch

# Wrong: size not in tuple (still works but less clear)
wrong_tensor = torch.zeros(2, 3)
print(wrong_tensor)

# Right: size as tuple
right_tensor = torch.zeros((2, 3))
print(right_tensor)

# Wrong: create tensor on GPU without checking
# Uncommenting the next line may cause error if no GPU
# gpu_tensor = torch.zeros((2, 3), device='cuda')

# Right: check before creating on GPU
if torch.cuda.is_available():
    gpu_tensor = torch.zeros((2, 3), device='cuda')
    print(gpu_tensor)
else:
    print('CUDA not available, skipping GPU tensor creation.')

Output

tensor([[0., 0., 0.],
        [0., 0., 0.]])
tensor([[0., 0., 0.],
        [0., 0., 0.]])
CUDA not available, skipping GPU tensor creation.

Quick Reference

Here is a quick summary of how to create zeros tensors in PyTorch:

Function	Description	Example
torch.zeros(size)	Creates a tensor of zeros with given shape	torch.zeros((2, 3))
torch.zeros(size, dtype=torch.int64)	Zeros tensor with integer type	torch.zeros((2, 3), dtype=torch.int64)
torch.zeros(size, device='cuda')	Zeros tensor on GPU device	torch.zeros((2, 3), device='cuda')
torch.zeros(size, requires_grad=True)	Zeros tensor tracked for gradients	torch.zeros((2, 3), requires_grad=True)

Key Takeaways

Use torch.zeros() with a size tuple to create a zeros tensor in PyTorch.

Specify dtype to control the data type of the zeros tensor.

Check for CUDA availability before creating tensors on GPU to avoid errors.

Passing size as a tuple is the clearest and recommended way.

Use requires_grad=True if you want to track operations for gradient computation.