Multi-Dimensional Array Slicing (Row, Column & Matrix Slices)0%
Memory Views vs Deep Copies (view() vs copy())

Multi-Dimensional Array Slicing (Row, Column & Matrix Slices)

Beginner12 min read•Updated: Jul 11, 2026
Study Materials

Multi-Dimensional Array Slicing (Row, Column & Matrix Slices)

In multi-dimensional arrays, slicing generalizes across multiple axes. Unlike standard Python (which requires clumsy chained brackets like matrix[row][col]), NumPy allows clean comma-separated slicing along every axis simultaneously: arr[row_slice, col_slice].

1. The 2D Coordinate Syntax: arr[rows, cols]

In a 2D matrix:

  • Axis 0 represents Rows (vertical dimension).
  • Axis 1 represents Columns (horizontal dimension).
Python
import numpy as np
 
# Create a 4x5 matrix (4 rows, 5 columns)
grid = np.array([
[10, 11, 12, 13, 14],
[20, 21, 22, 23, 24],
[30, 31, 32, 33, 34],
[40, 41, 42, 43, 44]
])
 
# Extract a single element at Row 2, Column 3
print(grid[2, 3]) # 33
Multi-Dimensional Array Slicing GridClick to Zoom
Multi-Dimensional Array Slicing Grid

2. Extracting Complete Rows and Columns

By using a lone colon : for an axis, you select all elements along that dimension:

Python
# Extract entire Row 1 (all columns)
row_1 = grid[1, :]
print(row_1) # [20 21 22 23 24]
 
# Extract entire Column 2 (all rows)
col_2 = grid[:, 2]
print(col_2) # [12 22 32 42]

3. Sub-Matrix Slicing (2D Blocks)

You can slice rows and columns simultaneously to extract sub-grids:

Python
# Extract Rows 1 to 3 (exclusive) and Columns 1 to 4 (exclusive)
sub_block = grid[1:3, 1:4]
print(sub_block)
# [[21 22 23]
# [31 32 33]]

4. Preserving Dimensions vs Reducing Dimensions

Notice the subtle difference between slicing with a single integer versus a slice range:

Python
# Case A: Dimensionality Reduction (Returns 1D array)
r1 = grid[0, :]
print(r1.shape) # (5,) -> ndim reduced from 2 to 1!
 
# Case B: Dimensionality Preserved (Returns 2D array with 1 row)
r2 = grid[0:1, :]
print(r2.shape) # (1, 5) -> ndim remains 2!

In machine learning algorithms and neural networks, keeping rank-2 shapes (1, 5) is often mandatory for matrix multiplication!

5. The Ellipsis Operator (...)

When slicing arrays with 4, 5, or more dimensions (such as video frames: [batch, frames, height, width, channels]), writing colons for every axis becomes tedious. The Ellipsis (...) automatically expands to as many full-slice colons : as needed:

Python
video = np.zeros((32, 100, 1080, 1920, 3))
 
# Select only the Red channel (channel 0) across all batches, frames, height, and width:
red_channel = video[..., 0] # Equivalent to video[:, :, :, :, 0]
print(red_channel.shape) # (32, 100, 1080, 1920)

Multiple Choice Questions

1. In a 2D NumPy array 'arr', how do you extract the entire third column across all rows?

A. arr[3, :] B. arr[:, 2] C. arr[2, :] D. arr[:, 3] Answer: B Explanation: Python uses 0-based indexing; column 3 is index 2. arr[:, 2] selects all rows (:) in column index 2.

2. How does 'arr[0, :]' differ from 'arr[0:1, :]'?

A. arr[0, :] returns a 1D vector of shape (N,); arr[0:1, :] preserves rank-2 dimensionality returning shape (1, N) B. They return completely different numbers C. arr[0:1, :] deletes the row D. There is no difference Answer: A Explanation: Indexing with a scalar integer reduces the dimension (rank), returning a 1D array; slicing with a range [0:1] retains the 2D matrix structure.

3. What does the Ellipsis operator '...' do in multidimensional slicing: 'tensor[..., 0]'?

A. Stops the execution B. Expands to fill as many complete slice colons (:) as necessary to span intermediate unspecified dimensions C. Fills the array with dots D. Generates random numbers Answer: B Explanation: The Ellipsis (...) represents all intervening dimensions not explicitly specified in the slice.

4. If matrix M has shape (6, 8), what will be the shape of 'M[2:5, 3:7]'?

A. (3, 4) B. (2, 3) C. (6, 8) D. (5, 7) Answer: A Explanation: Rows 2:5 yields 3 rows (indices 2, 3, 4). Columns 3:7 yields 4 columns (indices 3, 4, 5, 6). The resulting shape is (3, 4).

5. Why is 'matrix[1, 2]' preferred over 'matrix[1][2]' in NumPy?

A. matrix[1][2] creates a temporary intermediate array object for matrix[1], adding unnecessary overhead, whereas matrix[1, 2] accesses memory directly in one step B. matrix[1][2] is illegal syntax in Python C. matrix[1, 2] converts the number to string D. It saves the file to disk Answer: A Explanation: Chained indexing (matrix[1][2]) creates an unnecessary intermediate 1D slice object; multidimensional comma indexing (matrix[1, 2]) computes the memory offset directly in C.

Next Lesson

Memory Views vs Deep Copies (view() vs copy())

Continue learning with hands-on practice, examples, and exercises in the upcoming topic.

Related Lessons

Previous LessonNext Lesson
1D Array Indexing, Slicing & Step StridesMemory Views vs Deep Copies (view() vs copy())

Practice Quiz

Test your understanding of this lesson with 5 questions. Each question has one correct answer.

PreviousNext
PrevNext