🕯️LORE

Common Errors & Gotchas

Properties2
tagsnumpy, python, debugging, gotchas
aliasesValueError shapes, broadcast error

3 min read

Common Errors & Gotchas

Purpose

A troubleshooting reference for the errors that come up most often when working with NumPy, why they happen, and the fix.

ValueError: operands could not be broadcast together

a = np.array([1, 2, 3])          # shape (3,)
b = np.array([1, 2])                # shape (2,)
a + b                                  # ERROR — shapes incompatible

Debug checklist

Print .shape for both arrays. Broadcasting requires shapes to match or be 1 when compared from the right. Use arr[:, np.newaxis] or np.expand_dims() to explicitly align dimensions if needed.

View vs Copy Confusion (Silent Mutation)

sub = arr2d[0:2, 0:2]     # a VIEW, not independent
sub[0, 0] = 999              # also modifies arr2d!

Basic slicing returns a view; fancy/boolean indexing returns a copy

This asymmetry is the single most common source of “why did my original array change?” bugs. Use .copy() explicitly whenever independence is required. See 02-Indexing-Slicing.

Integer Division Surprise

np.array([1, 2, 3]) / np.array([2, 2, 2])     # [0.5, 1.0, 1.5] — true division, returns float64
np.array([1, 2, 3]) // np.array([2, 2, 2])       # [0, 1, 1] — floor division

In Python 3 / modern NumPy, / always performs true (float) division even between integer arrays — no need to cast to float first as in old Python 2 code.

dtype Upcasting Producing Unexpected Results

arr = np.array([1, 2, 3], dtype="int8")
arr + 200          # overflow — int8 max is 127, wraps around silently in some cases

Fixed-width integer types can silently overflow

Unlike Python’s arbitrary-precision int, NumPy’s int8/int16/int32 wrap around on overflow without raising an error by default. Check value ranges before choosing a narrow dtype, or use np.errstate to catch it: with np.errstate(over="raise"): ....

NaN Comparisons Always False

np.nan == np.nan          # False!
arr == np.nan                # never matches, even where NaN exists
np.isnan(arr)                  # correct way to check for NaN

Mixing and/or with Arrays

# BAD
if arr1 > 0 and arr2 > 0:
    ...
# ValueError: truth value of an array is ambiguous
 
# GOOD
if np.all((arr1 > 0) & (arr2 > 0)):
    ...

Python’s and/or/if expect a single boolean; use &/| for element-wise combination, and reduce with .any()/.all() before an if statement.

Reshape Errors — Element Count Mismatch

arr = np.arange(10)
arr.reshape(3, 4)      # ValueError: cannot reshape array of size 10 into shape (3,4)

New shape’s total element count must exactly match the original (3*4=12 ≠ 10). Use reshape(-1, n) to let NumPy infer one dimension automatically, or np.resize() if changing the total count is intentional.

Axis Confusion in Aggregations

arr2d.sum(axis=0)     # per-COLUMN sums (collapses rows)
arr2d.sum(axis=1)       # per-ROW sums (collapses columns)

If the result shape is the opposite of what you expected, you likely have axis=0/axis=1 swapped. See 04-Math-Statistical-Functions for the mental model.

Comparing Floats for Exact Equality

0.1 + 0.2 == 0.3            # False — floating-point precision
np.isclose(0.1 + 0.2, 0.3)    # True — use this instead
np.allclose(arr1, arr2)         # array-wide tolerance-based comparison

object dtype Sneaking In

np.array([1, 2, "three"])     # dtype becomes '<U21' (unicode string) — all elements coerced!
np.array([1, [2, 3], "x"])       # dtype becomes 'object' — loses vectorization entirely

Check .dtype after creating arrays from mixed or nested-irregular data — an unexpected object or string dtype silently disables fast vectorized math.

Graph View

Backlinks