How to Slice Tensor in PyTorch: Syntax and Examples

In PyTorch, you slice a tensor using standard Python slicing syntax inside square brackets, like tensor[start:stop:step]. You can slice along any dimension by specifying indices for each dimension separated by commas, for example, tensor[:, 1:4] slices all rows and columns 1 to 3.

Syntax

PyTorch tensor slicing uses Python's standard slice notation inside square brackets. You can specify start, stop, and step for each dimension separated by commas.

• tensor[start:stop:step] slices one dimension.
• tensor[dim1_slice, dim2_slice, ...] slices multiple dimensions.
• Omitting start means from the beginning, omitting stop means until the end, and omitting step means step of 1.

python

tensor[start:stop:step]
tensor[dim1_slice, dim2_slice, ...]

Example

This example shows how to slice a 2D tensor to get specific rows and columns using PyTorch slicing syntax.

python

import torch

# Create a 3x5 tensor with values from 0 to 14
tensor = torch.arange(15).reshape(3, 5)

# Slice: all rows, columns 1 to 3 (index 1 to 4 exclusive)
sliced = tensor[:, 1:4]

print("Original tensor:\n", tensor)
print("Sliced tensor (all rows, cols 1 to 3):\n", sliced)

Output

Original tensor:
 tensor([[ 0,  1,  2,  3,  4],
        [ 5,  6,  7,  8,  9],
        [10, 11, 12, 13, 14]])
Sliced tensor (all rows, cols 1 to 3):
 tensor([[ 1,  2,  3],
        [ 6,  7,  8],
        [11, 12, 13]])

Common Pitfalls

Common mistakes when slicing tensors include:

• Using out-of-range indices causing empty slices or errors.
• Confusing inclusive vs exclusive end index (Python slices exclude the stop index).
• Forgetting to use commas to separate slices for multiple dimensions.
• Modifying slices expecting original tensor to change (slices are views, but some operations create copies).

python

import torch

tensor = torch.arange(10)

# Wrong: missing comma for multi-dim slice (will error if tensor is 2D)
# sliced_wrong = tensor[1:5 2:7]  # SyntaxError

# Right: use comma to separate dimensions
# For 2D tensor example:
tensor_2d = torch.arange(20).reshape(4,5)
sliced_right = tensor_2d[1:3, 2:5]

print(sliced_right)

Output

tensor([[ 7,  8,  9],
        [12, 13, 14]])

Quick Reference

Syntax	Description
tensor[start:stop]	Slice from start index up to but not including stop index
tensor[start:stop:step]	Slice with step size (e.g., every 2nd element)
tensor[:, col_start:col_stop]	Slice all rows and specific columns
tensor[row_start:row_stop, :]	Slice specific rows and all columns
tensor[dim1_slice, dim2_slice, ...]	Slice multiple dimensions with comma-separated slices

Key Takeaways

Use Python slice syntax inside square brackets to slice PyTorch tensors.

Separate slices for each dimension with commas to slice multi-dimensional tensors.

Remember Python slices exclude the stop index, so stop is not included in the slice.

Slicing returns a view, so changes to the slice may affect the original tensor.

Avoid out-of-range indices to prevent empty slices or errors.