np.stack: Joining on a New Axis

You have the squad's scores for 2023 in one (2, 3) table and for 2024 in another (2, 3) table. You don't want a longer table — you want a 3D array where the first index picks the season: seasons[0] is 2023, seasons[1] is 2024. That is exactly what np.stack does.

Definition

np.stack joins a sequence of arrays of exactly the same shape along a new axis. The result has one more dimension than the inputs, and the new axis has length equal to the number of arrays stacked.

• Same shape — every input must match exactly; no size can differ.
• New axis — concatenate reuses an existing axis; stack inserts a fresh one. Two (2, 3) arrays → (2, 2, 3).
• axis argument — chooses where the new axis goes in the result shape (0 = front, -1 = end).

Syntax

np.stack(arrays, axis=0, out=None, *, dtype=None, casting="same_kind")

Example — 1D: stack two players

import numpy as np

virat = np.array([72, 45, 88, 31])
rohit = np.array([56,  0, 74, 88])

s0 = np.stack((virat, rohit))
print(s0)
print(s0.shape)
s1 = np.stack((virat, rohit), axis=1)
print(s1)
print(s1.shape)
print(np.concatenate((virat, rohit)).shape)

[[72 45 88 31]
 [56  0 74 88]]
(2, 4)
[[72 56]
 [45  0]
 [88 74]
 [31 88]]
(4, 2)
(8,)

How it works: two (4,) arrays, so the result is 2D.

• axis=0 → new axis in front → (2, 4): same as vstack.
• axis=1 → new axis at the end → (4, 2): same as column_stack.
• concatenate stays 1D → (8,).

Example — 2D: stack seasons into a 3D array

s2023 = np.array([[72, 45, 88],
                  [56,  0, 74]])
s2024 = np.array([[31, 64,  0],
                  [88, 32, 45]])
seasons = np.stack((s2023, s2024))
print(seasons.shape)
print(seasons)
print(seasons.sum(axis=0))

(2, 2, 3)
[[[72 45 88]
  [56  0 74]]

 [[31 64  0]
  [88 32 45]]]
[[103 109  88]
 [144  32 119]]

How it works: shape (2, 2, 3) = (season, player, match). seasons[1, 0, 2] is Virat's 3rd match of 2024. Summing over axis=0 (the new "season" axis) adds the two seasons element by element.

Where does the new axis go?

print(np.stack((s2023, s2024), axis=2).shape)
print(np.stack((s2023, s2024), axis=-1)[0])
print(np.concatenate((s2023, s2024)).shape)

(2, 3, 2)
[[72 31]
 [45 64]
 [88  0]]
(4, 3)

With axis=-1 the season becomes the last axis: result[0] is Virat, and each row pairs his 2023 and 2024 score for that match.

stack vs concatenate

Tip: np.stack is how you turn a Python list of same-shape arrays into one array — e.g. a list of 100 image arrays (28, 28) → np.stack(images) gives (100, 28, 28).

Common mistakes

• ❌ Stacking arrays of different shapes. ✅ np.stack raises ValueError: all input arrays must have the same shape — use concatenate instead if you want to join different lengths.
• ❌ Using stack when you only wanted more rows. ✅ concatenate/vstack keep 2D; stack makes it 3D.
• ❌ Forgetting the new axis length equals the number of arrays, not their size.

Interview tip

Q: You have 3 RGB images, each (64, 64, 3). What do np.stack(imgs) and np.concatenate(imgs) return? A: stack → (3, 64, 64, 3), a batch of 3 images. concatenate → (192, 64, 3), the images glued vertically into one tall image.