Random numbers with Generator

The modern way: `default_rng`

You create a generator with a seed (seed) and request numbers from it. The same seed produces always the same sequence, and that is what makes a result reproducible:

import numpy as np

rng = np.random.default_rng(42)
print(rng.random(3))
print(rng.integers(0, 10, size=5))
print(rng.normal(loc=0, scale=1, size=3))
print(rng.choice([10, 20, 30], size=2, replace=False))

[0.77395605 0.43887844 0.85859792]
[0 6 2 0 5]
[ 0.1278404  -0.31624259 -0.01680116]
[30 20]

a = np.random.default_rng(7).random(3)
b = np.random.default_rng(7).random(3)
print(np.array_equal(a, b))

True

Why not `np.random.seed`

The old API (np.random.seed(42) followed by np.random.rand(3)) still works and appears in a lot of material, including courses, but uses a global state hidden in the module:

np.random.seed(42)
print(np.random.rand(3))

[0.37454012 0.95071431 0.73199394]

The problem with global state is that any code, anywhere, can consume numbers from it and change its sequence without you noticing, which breaks reproducibility and makes tests fragile. With the Generator, the state resides in a controlled object. And, therefore, functions that use randomness must receive the generator as a parameter:

def sorteio(n, rng):
    return rng.integers(1, 7, size=n)


print(sorteio(5, np.random.default_rng(1)).tolist())

[3, 4, 5, 6, 1]

With this signature, the function test passes a generator with a fixed seed and checks for an exact result, while production passes any generator.

The guarantee of reproduction

The same seed reproduces the same numbers in the same version of NumPy. Between versions, the guarantee of the Generator is weaker than that of the old API, which the project froze. For an experiment that needs to be reproduced years later, also record the version of the library.

Shuffle: in place or in copy

The rng.shuffle shuffles the array itself and returns None. Writing x = rng.shuffle(x) discards the array. To get a shuffled version and keep the original, use permutation:

vetor = np.arange(1, 6)
resultado = rng.shuffle(vetor)
print(resultado, sorted(vetor.tolist()))

original = np.arange(1, 6)
embaralhado = rng.permutation(original)
print(original, sorted(embaralhado.tolist()))

None [1, 2, 3, 4, 5]
[1 2 3 4 5] [1, 2, 3, 4, 5]

Generate a synthetic dataset

Synthetic data with known characteristics allows testing an analysis knowing the answer. Here, a thousand heights with a mean of 170 and a standard deviation of 10: the sample mean is close to 170, but not exactly at it, because it is a sample:

alturas = rng.normal(170, 10, size=1000)
print(abs(alturas.mean() - 170) < 1, abs(alturas.std() - 10) < 1)

sorteados = rng.choice(["cara", "coroa"], size=1000, p=[0.9, 0.1])
print((sorteados == "cara").mean() > 0.8)

True True
True