How to Squeeze Tensor in PyTorch: Syntax and Examples

In PyTorch, use torch.squeeze(tensor) to remove all dimensions of size 1 from a tensor. You can also specify a dimension with torch.squeeze(tensor, dim) to remove only that dimension if its size is 1.

Syntax

The torch.squeeze function removes dimensions of size 1 from a tensor. It has two forms:

• torch.squeeze(input): removes all dimensions with size 1.
• torch.squeeze(input, dim): removes the dimension dim if its size is 1; otherwise, returns the tensor unchanged.

This helps simplify tensor shapes by eliminating unnecessary dimensions.

python

torch.squeeze(input, dim=None)

Example

This example shows how to use torch.squeeze to remove all size-1 dimensions and how to remove a specific dimension.

python

import torch

# Create a tensor with shape (1, 3, 1, 5)
t = torch.randn(1, 3, 1, 5)
print(f"Original shape: {t.shape}")

# Remove all dimensions of size 1
t_squeezed = torch.squeeze(t)
print(f"Shape after squeeze(): {t_squeezed.shape}")

# Remove only dimension 2 (which has size 1)
t_squeezed_dim = torch.squeeze(t, dim=2)
print(f"Shape after squeeze(dim=2): {t_squeezed_dim.shape}")

# Trying to squeeze a dimension that is not size 1 returns the original tensor
no_change = torch.squeeze(t, dim=1)
print(f"Shape after squeeze(dim=1) (no change): {no_change.shape}")

Output

Original shape: torch.Size([1, 3, 1, 5])
Shape after squeeze(): torch.Size([3, 5])
Shape after squeeze(dim=2): torch.Size([1, 3, 5])
Shape after squeeze(dim=1) (no change): torch.Size([1, 3, 1, 5])

Common Pitfalls

Common mistakes when using torch.squeeze include:

• Trying to squeeze a dimension that is not size 1, which does nothing and may confuse you.
• Assuming squeeze() changes the original tensor; it returns a new tensor instead.
• Not checking tensor shape before squeezing, which can lead to unexpected shape changes.

Always verify the tensor shape before and after squeezing.

python

import torch

t = torch.randn(2, 3, 1)
print(f"Original shape: {t.shape}")

# Wrong: trying to squeeze dim=0 which is size 2
squeezed_wrong = torch.squeeze(t, dim=0)
print(f"Shape after wrong squeeze(dim=0): {squeezed_wrong.shape}")

# Right: squeeze dim=2 which is size 1
squeezed_right = torch.squeeze(t, dim=2)
print(f"Shape after right squeeze(dim=2): {squeezed_right.shape}")

Output

Original shape: torch.Size([2, 3, 1])
Shape after wrong squeeze(dim=0): torch.Size([2, 3, 1])
Shape after right squeeze(dim=2): torch.Size([2, 3])

Quick Reference

Tips for using torch.squeeze:

• Use torch.squeeze(tensor) to remove all size-1 dimensions.
• Use torch.squeeze(tensor, dim) to remove a specific size-1 dimension.
• If the specified dimension is not size 1, the tensor stays the same.
• Check tensor shapes before and after squeezing to avoid shape errors.

Key Takeaways

Use torch.squeeze(tensor) to remove all dimensions of size 1 from a tensor.

Specify dim in torch.squeeze(tensor, dim) to remove only that dimension if its size is 1.

Squeezing a dimension that is not size 1 does not change the tensor shape.

torch.squeeze returns a new tensor; it does not modify the original tensor in place.

Always check tensor shapes before and after squeezing to avoid unexpected results.