Indexing, Slicing and Reshaping

Topic — Getting at parts of an array, and changing its shape. The one thing that genuinely differs from Python lists is in the middle of this page: slices are views.

---

1-D indexing and slicing

Identical to strings and lists — 0 first, -1 last, stop excluded.

arr = np.array([10, 20, 30, 40, 50])
print(arr[0])       # → 10
print(arr[2])       # → 30
print(arr[-1])      # → 50
print(arr[1:4])     # → [20 30 40]
print(arr[:3])      # → [10 20 30]
print(arr[2:])      # → [30 40 50]
print(arr[::2])     # → [10 30 50]
print(arr[::-1])    # → [50 40 30 20 10]

Output

10
30
50
[20 30 40]
[10 20 30]
[30 40 50]
[10 30 50]
[50 40 30 20 10]

Nothing new so far. That's the point — the syntax carries over.

---

2-D indexing

Here NumPy adds something lists don't have: one bracket, comma-separated per dimension.

a = np.array([[1, 2, 3],
              [4, 5, 6],
              [7, 8, 9]])

print(a[1, 2])    # → 6    row 1, column 2
print(a[1][2])    # → 6    same answer, worse way

Output

6
6

a[1, 2] reads "row 1, column 2" in one operation. a[1][2] builds an intermediate array for row 1, then indexes into it — same answer, more work. Prefer the comma form.

The mental model:

        col 0  col 1  col 2
row 0 [   1      2      3  ]
row 1 [   4      5      6  ]
row 2 [   7      8      9  ]

Rows and columns

print(a[0])        # → [1 2 3]   whole row
print(a[:, 0])     # → [1 4 7]   whole column

Output

[1 2 3]
[1 4 7]

a[0] is row 0. For a column you need a[:, 0] — "all rows, column 0". Extracting a column from a list of lists takes a comprehension; here it's punctuation.

2-D slicing

Slice each dimension independently:

print(a[1, 2])        # → 6                single element
print(a[:2, :2])      # first 2 rows, first 2 columns
print(a[1:, :2])      # from row 1, first 2 columns
print(a[:, 1])        # → [2 5 8]          column 1
print(a[::2, ::2])    # every other row and column

Output

6
[[1 2]
 [4 5]]
[[4 5]
 [7 8]]
[2 5 8]
[[1 3]
 [7 9]]

Indexing a dimension drops it; slicing keeps it

print(a[:, 1].shape)      # → (3,)    1-D — the column dimension is gone
print(a[:, 1:2].shape)    # → (3, 1)  still 2-D — a one-column matrix

a[:, 1] uses an index for the column, so that dimension collapses. a[:, 1:2] uses a slice, so it survives with length 1. This bites when a function insists on 2-D input.

---

Slices are views, not copies

This is the most important difference between NumPy and everything else you've seen.

s = np.array([1, 2, 3, 4, 5])
sl = s[1:3]
sl[0] = 99
print(s)   # → [ 1 99  3  4  5]

Output

[ 1 99  3  4  5]

Writing to the slice changed the original. Compare a Python list, which copies:

L = [1, 2, 3, 4, 5]
Ls = L[1:3]
Ls[0] = 99
print(L)   # → [1, 2, 3, 4, 5]   untouched

Output

[1, 2, 3, 4, 5]

This is the reverse of list behaviour

On the Lists page, lst[:] is the idiomatic way to copy a list. On a NumPy array, arr[:] is a view into the same memory — the opposite. Carrying the list habit over to NumPy produces silent data corruption.

A slice is deliberately a view: NumPy arrays are often large, and copying a million-row slice just to read it would be wasteful. The trade is that writes propagate.

Checking, and opting out

sl = s[1:3]
print(sl.base is not None)   # → True   this array borrows someone else's memory

c = s[1:3].copy()
c[0] = 777
print(s)   # → unchanged by the write to c

Output

True

.base is None for an array that owns its data, and points at the original for a view. Call .copy() whenever you intend to modify a slice independently.

Operation	Returns
arr[1:3]	view
arr[:, 0]	view
arr.reshape(...)	view
arr.ravel()	view (usually)
arr.T	view
arr.copy()	copy
arr.flatten()	copy
arr[[0, 2]] (fancy)	copy
arr[arr > 5] (boolean)	copy

The pattern: slicing and reshaping give views; selecting arbitrary elements gives copies. A view has to be describable as "start here, step by this much", and an arbitrary selection isn't.

---

Boolean masking

Comparing an array gives an array of booleans, which you can then use as an index.

a = np.array([10, 20, 30, 40])
print(a > 20)        # → [False False  True  True]   the mask
print(a[a > 20])     # → [30 40]                     the selection

Output

[False False  True  True]
[30 40]

Read a[a > 20] as "the elements of a where a > 20". This replaces a filtering loop entirely.

Combining conditions

print(a[(a > 15) & (a < 40)])   # → [20 30]
print(a[~(a > 20)])             # → [10 20]
print((a > 20).sum())           # → 2

Output

[20 30]
[10 20]
2

	Use	Not
and	&	and
or	|	or
not	~	not

warning

and / or / not don't work on arrays

They ask for one True/False, and an array of many values can't answer. You get ValueError: The truth value of an array with more than one element is ambiguous. Use &, |, ~ — and parenthesise each condition, because & binds tighter than >.

(a > 20).sum() counts matches, because True is 1 — the same boolean-as-int fact from the basics.

np.where

Two jobs, depending on argument count.

a = np.array([10, 20, 30, 40])
print(np.where(a > 20, a, 0))   # → [ 0  0 30 40]   choose per element
print(np.where(a > 20))         # → (array([2, 3]),)  the indices
print(a[np.where(a > 20)])      # → [30 40]

Output

[ 0  0 30 40]
(array([2, 3]),)
[30 40]

Three arguments is a vectorised if/else — "where the condition holds take a, otherwise take 0". One argument returns the positions where it holds, as a tuple of index arrays (one per dimension, hence the trailing comma).

Fancy indexing

Index with a list of positions, in any order, with repeats:

a = np.array([10, 20, 30, 40, 50])
print(a[[0, 2, 4]])    # → [10 30 50]
print(a[[4, 4, 0]])    # → [50 50 10]

Output

[10 30 50]
[50 50 10]

The result is a copy, and its length is the length of your index list — not the array's.

---

Reshaping

reshape reorganises the same elements into a new shape.

b = np.arange(1, 7)
print(b)               # → [1 2 3 4 5 6]
print(b.reshape(2, 3))
print(b.reshape(3, 2))
print(b.reshape(6, 1))

Output

[1 2 3 4 5 6]
[[1 2 3]
 [4 5 6]]
[[1 2]
 [3 4]
 [5 6]]
[[1]
 [2]
 [3]
 [4]
 [5]
 [6]]

Elements fill row by row.

The legality rule

b.reshape(4, 2)

Error

ValueError: cannot reshape array of size 6 into shape (4,2)

The product of the new shape must equal .size. Six elements reshape to (2,3), (3,2), (6,1), (1,6) — because each multiplies to 6. (4,2) needs 8 elements, so there's nothing to put in the last two slots and NumPy refuses rather than inventing values.

Check with arr.size before reshaping, or let NumPy do the arithmetic:

print(b.reshape(-1, 2))

Output

[[1 2]
 [3 4]
 [5 6]]

-1 means "work this dimension out from the others". Only one -1 is allowed.

Reshape returns a view

r = b.reshape(2, 3)
r[0, 0] = 99
print(b)   # → [99  2  3  4  5  6]

Output

[99  2  3  4  5  6]

Same trap as slicing — reshape doesn't copy. If you need an independent result, chain .copy().

Flattening

Both collapse to 1-D; the difference is the view/copy question again.

m = np.array([[1, 2], [3, 4]])

f = m.flatten()
f[0] = 99
print(m.tolist())   # → [[1, 2], [3, 4]]     flatten COPIES

rv = m.ravel()
rv[0] = 99
print(m.tolist())   # → [[99, 2], [3, 4]]    ravel VIEWS

Output

[[1, 2], [3, 4]]
[[99, 2], [3, 4]]

flatten() is the safe one. ravel() is faster because it avoids the copy when it can.

Transpose

.T swaps the axes — rows become columns.

m = np.array([[1, 2, 3],
              [4, 5, 6]])
print(m.shape, "->", m.T.shape)   # → (2, 3) -> (3, 2)
print(m.T)

Output

(2, 3) -> (3, 2)
[[1 4]
 [2 5]
 [3 6]]

Also a view — and note it's .T, an attribute, with no parentheses.

---

Common Mistakes

#	Mistake	Wrong	Right
1	Assuming a slice copies	arr[1:3] shares memory	arr[1:3].copy()
2	arr[:] to copy	it's a view	arr.copy()
3	Assuming reshape copies	it's a view	arr.reshape(2,3).copy()
4	Chained 2-D indexing	a[1][2]	a[1, 2]
5	and in a mask	ValueError: truth value ... ambiguous	&
6	Unparenthesised mask	a > 15 & a < 40	(a > 15) & (a < 40)
7	Reshape to a wrong size	size 6 into shape (4,2)	product must equal .size
8	.T() with parentheses	TypeError	.T

---

Summary

Task	Syntax
Element (2-D)	a[row, col]
Whole row / column	a[i] / a[:, j]
Sub-block	a[r1:r2, c1:c2]
Filter by condition	a[a > x]
Vectorised if/else	np.where(cond, x, y)
Positions matching	np.where(cond)
Pick specific indices	a[[i, j, k]]
Change shape	.reshape(r, c) or .reshape(-1, c)
Collapse to 1-D	.flatten() (copy) / .ravel() (view)
Swap axes	.T
Force a copy	.copy()

Key takeaways

• 1-D indexing and slicing match lists and strings exactly
• 2-D uses one bracket with a comma: a[1, 2], and a[:, 0] for a column
• Indexing a dimension drops it; slicing keeps it — (3,) vs (3, 1)
• Slices, reshapes, ravel and .T are views — writing to them writes to the original
• Boolean masks and fancy indexing return copies
• Use &, |, ~ in masks, never and/or/not, and parenthesise each condition
• reshape is legal only when the new shape's product equals .size; -1 infers one dimension
• .copy() is the escape hatch whenever you plan to modify

See also: Arrays and Attributes · Operations and Broadcasting · Strings and Slicing for the syntax this builds on

---

Run It Yourself

numpy_indexing.py

import numpy as np

a = np.array([[1, 2, 3],
              [4, 5, 6],
              [7, 8, 9]])

print(a)
print(f"{'a[1, 2]':<14} {a[1, 2]}")
print(f"{'row a[0]':<14} {a[0]}")
print(f"{'col a[:, 0]':<14} {a[:, 0]}")
print(f"{'a[:2, :2]':<14} {a[:2, :2].tolist()}")
print(f"{'a[::2, ::2]':<14} {a[::2, ::2].tolist()}")

# masking
flat = a.ravel()
print(f"{'flat':<14} {flat}")
print(f"{'mask > 5':<14} {flat > 5}")
print(f"{'flat[> 5]':<14} {flat[flat > 5]}")
print(f"{'count > 5':<14} {(flat > 5).sum()}")
print(f"{'where':<14} {np.where(flat > 5, flat, 0)}")

# the view trap, demonstrated
original = np.array([1, 2, 3, 4, 5])
view = original[1:3]
view[0] = 99
print(f"{'after view':<14} {original}")

original = np.array([1, 2, 3, 4, 5])
copy = original[1:3].copy()
copy[0] = 99
print(f"{'after copy':<14} {original}")

Output

[[1 2 3]
 [4 5 6]
 [7 8 9]]
a[1, 2]        6
row a[0]       [1 2 3]
col a[:, 0]    [1 4 7]
a[:2, :2]      [[1, 2], [4, 5]]
a[::2, ::2]    [[1, 3], [7, 9]]
flat           [1 2 3 4 5 6 7 8 9]
mask > 5       [False False False False False  True  True  True  True]
flat[> 5]      [6 7 8 9]
count > 5      4
where          [0 0 0 0 0 6 7 8 9]
after view     [ 1 99  3  4  5]
after copy     [1 2 3 4 5]

---

Practice Questions

From the Unit 1 question bank. Tags and marks are explained on the Python index.

Unit 1 § J — NumPy

J5. [OUT] 2-D slicing. [5]

import numpy as np
a = np.array([[1, 2, 3],
              [4, 5, 6],
              [7, 8, 9]])

print(a[1, 2])
print(a[:2, :2])
print(a[1:, :2])
print(a[:, 1])
print(a[::2, ::2])

J7. [OUT] One of these raises an error. Which, and what is the error message? [3]

import numpy as np
a = np.array([1, 2, 3, 4, 5, 6])
print(a.reshape(2, 3))
print(a.reshape(3, 2))
print(a.reshape(6, 1))
print(a.reshape(4, 2))

State the rule that decides whether a reshape is legal.