How to Use view() in PyTorch for Tensor Reshaping

In PyTorch, view() is used to reshape a tensor without changing its data. You provide the new shape as arguments, and one dimension can be set to -1 to infer its size automatically.

Syntax

The view() function changes the shape of a tensor. You call it on a tensor and pass the new shape as arguments. Use -1 for one dimension to let PyTorch calculate it automatically.

• tensor.view(new_shape): Returns a tensor with the specified shape.
• new_shape: A tuple or list of integers representing the desired shape.
• -1: Automatically infers the size of this dimension.

python

tensor.view(shape)
# Example:
tensor.view(2, 3)  # reshape tensor to 2 rows and 3 columns

Example

This example shows how to reshape a 1D tensor of 6 elements into a 2D tensor with 2 rows and 3 columns using view(). It also demonstrates using -1 to automatically infer one dimension.

python

import torch

# Create a 1D tensor with 6 elements
x = torch.arange(6)
print('Original tensor:', x)

# Reshape to 2 rows and 3 columns
x_reshaped = x.view(2, 3)
print('Reshaped tensor (2,3):\n', x_reshaped)

# Use -1 to infer the second dimension automatically
x_reshaped_auto = x.view(2, -1)
print('Reshaped tensor with -1 (2, -1):\n', x_reshaped_auto)

Output

Original tensor: tensor([0, 1, 2, 3, 4, 5])
Reshaped tensor (2,3):
 tensor([[0, 1, 2],
        [3, 4, 5]])
Reshaped tensor with -1 (2, -1):
 tensor([[0, 1, 2],
        [3, 4, 5]])

Common Pitfalls

Common mistakes when using view() include:

• Trying to reshape a tensor to a shape that does not match the total number of elements.
• Using more than one -1 in the shape, which is not allowed.
• Not ensuring the tensor is contiguous in memory before calling view(), which can cause errors.

Always check the total elements match and use tensor.contiguous() if needed.

python

import torch

x = torch.arange(6)

# Wrong: shape does not match total elements
try:
    x.view(4, 2)
except RuntimeError as e:
    print('Error:', e)

# Wrong: more than one -1
try:
    x.view(-1, -1)
except RuntimeError as e:
    print('Error:', e)

# Correct: make tensor contiguous before view if needed
x_t = x.t()  # transpose makes non-contiguous tensor
try:
    x_t.view(6)
except RuntimeError as e:
    print('Error:', e)

x_t_contig = x_t.contiguous()
print('Contiguous tensor view:', x_t_contig.view(6))

Output

Error: shape '[4, 2]' is invalid for input of size 6
Error: only one dimension can be inferred
Error: view size is not compatible with input tensor's size and stride (at least one dimension spans across two contiguous subspaces)
Contiguous tensor view: tensor([0, 1, 2, 3, 4, 5])

Quick Reference

Usage	Description
tensor.view(shape)	Reshape tensor to given shape
-1 in shape	Automatically infer dimension size
Only one -1 allowed	Multiple -1s cause error
tensor.contiguous()	Make tensor memory contiguous before view
Total elements must match	Shape must fit total number of elements

Key Takeaways

Use tensor.view(new_shape) to reshape tensors without changing data.

Only one dimension can be set to -1 to infer its size automatically.

Ensure the total number of elements matches the new shape.

Call tensor.contiguous() if the tensor is not contiguous before using view().

Common errors come from incompatible shapes or multiple -1s.