A NumPy Banana Skin • `dtype` (A NumPy for Numpties article)

A peculiar article about a peculiar issue with NumPy and an array's `dtype`

Today’s “article” is a bit different from my usual ones. Quite a bit different, it’s fair to say. Let me know what you think. Normal service resumes from the next article (if there’s such a thing as normal with my articles.)

Code shown in both formats: as a formatted image and as a native Substack code block

Here’s the same code in Substack’s native code blocks (but don’t forget to scroll down to see the article’s cover image!):

>>> import numpy as np

# I played a game every weekday last week
>>> n_points_earned_per_day = np.array(
...     [5, 4, 9, 2, 6]
... )

# Weird rule: points earned on first day are only worth half
>>> n_points_earned_per_day[0] = n_points_earned_per_day[0] / 2

>>> n_points_earned_per_day
array([2, 4, 9, 2, 6])

# Hmmm… Day 1 doesn't look right
# (Just checking my arithmetic…)
>>> 5 / 2
2.5

>>> n_points_earned_per_day
array([2, 4, 9, 2, 6])

>>> n_points_earned_per_day[0]
2

# ????

>>> n_points_earned_per_day.dtype
dtype('int64')

# Oh, I see…
>>> np.int64(5 / 2)
2

# OK

>>> n_points_earned_per_day = np.array(
...     [5, 4, 9, 2, 6],
...     dtype=np.float64,
... )

# Let's try again with the first day points adjustment
>>> n_points_earned_per_day[0] = n_points_earned_per_day[0] / 2

>>> n_points_earned_per_day
array([2.5, 4. , 9. , 2. , 6. ])

# That's better
# All good now
# Phew!
#
# (I'm glad I caught that before it caused any problems)
# (I'll be more careful next time)

#####
# I'm addicted to the game
# I started keeping scores in a NumPy array

>>> one_week_scores = np.array(
...     ["Stephen", "2024-02-26", 5, 4, 9, 2, 6]
... )

# New rule: points earned on last day are doubled
>>> one_week_scores[-1] = one_week_scores[-1] * 2

>>> one_week_scores
array(['Stephen', '2024-02-26', '5', '4', '9', '2', '66'], dtype='<U21')

# Eh?!
>>> one_week_scores[-1]
'66'

>>> one_week_scores.dtype
dtype('<U21')
# (That's NumPy-speak for "Unicode string of max 21 chars")

# Maybe this??
>>> one_week_scores = np.array(
...     ["Stephen", "2024-02-26", 5, 4, 9, 2, 6],
...     dtype=int,
... )
Traceback (most recent call last):
 ...
ValueError: invalid literal for int() with base 10: 'Stephen'

# Can only have one dtype in a NumPy array

# Back to the drawing board
>>> one_week_scores = np.array(
...     ["Stephen", "2024-02-26", 5, 4, 9, 2, 6],
... )

>>> scores_only = one_week_scores[2:].astype(int)
>>> scores_only
array([5, 4, 9, 2, 6])

# Last day rule
>>> scores_only[-1] = scores_only[-1] * 2

>>> scores_only
array([ 5,  4,  9,  2, 12])

# Much better

Code in this article uses Python 3.12