python-mastery

# theory

the slow way

You can iterate over a NumPy array with a Python for loop. It works. It's also one to two orders of magnitude slower than the vectorized version.

# Slow: Python loop runs once per element
result = []
for x in a:
    result.append(x ** 2 + 1)
result = np.array(result)

# Fast: one C-level call to do the same thing
result = a ** 2 + 1

The vectorized form is also shorter, easier to read, and impossible to get wrong on the indexing.

vectorized conditionals

np.where is the vectorized if/else. Branchless, no Python loop.

# Slow:
out = []
for x in a:
    out.append(0 if x < 0 else x)
out = np.array(out)

# Fast:
out = np.where(a < 0, 0, a)

For more than two branches, np.select:

np.select(
    [a < 0, a < 100, a >= 100],
    ["below_zero", "small", "big"],
    default="unknown",
)

import numpy as np
import time

# 100,000 random values
rng = np.random.default_rng(0)
a = rng.integers(low=-50, high=50, size=100_000)

# Slow: pure Python loop
t0 = time.perf_counter()
loop_result = np.zeros_like(a)
for i, x in enumerate(a):
    loop_result[i] = 0 if x < 0 else x * 2
loop_secs = time.perf_counter() - t0

# Fast: vectorized with np.where
t0 = time.perf_counter()
vec_result = np.where(a < 0, 0, a * 2)
vec_secs = time.perf_counter() - t0

print(f"loop:       {loop_secs*1000:7.1f} ms")
print(f"vectorized: {vec_secs*1000:7.1f} ms")
print(f"speedup:    {loop_secs / vec_secs:.0f}x")
print("results match:", np.array_equal(loop_result, vec_result))


# Given temps = np.array([-5, 12, -3, 20, 0, 8, -10, 15]), replace every negative value with 0 using np.where (no loops). Print 'positive_only: [...]' showing the resulting array.
# Your code here:

🐍

Loading PythonSetting up pandas & numpy...

# challenge 02/02todo

Categorize purchase amounts = np.array([3, 18, 55, 130, 240, 800]) into 'small' (<20), 'medium' (<100), 'large' (<500), or 'enterprise' (>=500). Print 'tiers:' followed by the resulting array on the next line.

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

import numpy as np
import time

# 100,000 random values
rng = np.random.default_rng(0)
a = rng.integers(low=-50, high=50, size=100_000)

# Slow: pure Python loop
t0 = time.perf_counter()
loop_result = np.zeros_like(a)
for i, x in enumerate(a):
    loop_result[i] = 0 if x < 0 else x * 2
loop_secs = time.perf_counter() - t0

# Fast: vectorized with np.where
t0 = time.perf_counter()
vec_result = np.where(a < 0, 0, a * 2)
vec_secs = time.perf_counter() - t0

print(f"loop:       {loop_secs*1000:7.1f} ms")
print(f"vectorized: {vec_secs*1000:7.1f} ms")
print(f"speedup:    {loop_secs / vec_secs:.0f}x")
print("results match:", np.array_equal(loop_result, vec_result))


# Categorize purchase amounts = np.array([3, 18, 55, 130, 240, 800]) into 'small' (<20), 'medium' (<100), 'large' (<500), or 'enterprise' (>=500). Print 'tiers:' followed by the resulting array on the next line.
# Your code here:

🐍

Loading PythonSetting up pandas & numpy...

Vectorization vs loops

the slow way

vectorized conditionals

See also