What’s new or different#

NumPy 1.17.0 introduced Generator as an improved replacement for the legacy RandomState. Here is a quick comparison of the two implementations.

Feature	Older Equivalent	Notes
`Generator`	`RandomState`	`Generator` requires a stream source, called a `BitGenerator` A number of these are provided. `RandomState` uses the Mersenne Twister `MT19937` by default, but can also be instantiated with any BitGenerator.
`random`	`random_sample`, `rand`	Access the values in a BitGenerator, convert them to `float64` in the interval `[0.0., 1.0)`. In addition to the `size` kwarg, now supports `dtype='d'` or `dtype='f'`, and an `out` kwarg to fill a user-supplied array. Many other distributions are also supported.
`integers`	`randint`, `random_integers`	Use the `endpoint` kwarg to adjust the inclusion or exclusion of the `high` interval endpoint.

The normal, exponential and gamma generators use 256-step Ziggurat methods which are 2-10 times faster than NumPy’s default implementation in standard_normal, standard_exponential or standard_gamma. Because of the change in algorithms, it is not possible to reproduce the exact random values using Generator for these distributions or any distribution method that relies on them.

In [1]: import numpy.random

In [2]: rng = np.random.default_rng()

In [3]: %timeit -n 1 rng.standard_normal(100000)
   ...: %timeit -n 1 numpy.random.standard_normal(100000)
   ...: 
1.45 ms +- 42 us per loop (mean +- std. dev. of 7 runs, 1 loop each)
2.77 ms +- 72.7 us per loop (mean +- std. dev. of 7 runs, 1 loop each)

In [4]: %timeit -n 1 rng.standard_exponential(100000)
   ...: %timeit -n 1 numpy.random.standard_exponential(100000)
   ...: 
673 us +- 63.6 us per loop (mean +- std. dev. of 7 runs, 1 loop each)
2.07 ms +- 29 us per loop (mean +- std. dev. of 7 runs, 1 loop each)

In [5]: %timeit -n 1 rng.standard_gamma(3.0, 100000)
   ...: %timeit -n 1 numpy.random.standard_gamma(3.0, 100000)
   ...: 
2.68 ms +- 88.2 us per loop (mean +- std. dev. of 7 runs, 1 loop each)
5.71 ms +- 55.8 us per loop (mean +- std. dev. of 7 runs, 1 loop each)

integers is now the canonical way to generate integer random numbers from a discrete uniform distribution. This replaces both randint and the deprecated random_integers.
The rand and randn methods are only available through the legacy RandomState.
Generator.random is now the canonical way to generate floating-point random numbers, which replaces RandomState.random_sample, sample, and ranf, all of which were aliases. This is consistent with Python’s random.random.
All bit generators can produce doubles, uint64s and uint32s via CTypes (ctypes) and CFFI (cffi). This allows these bit generators to be used in numba.
The bit generators can be used in downstream projects via Cython.
All bit generators use SeedSequence to convert seed integers to initialized states.
Optional dtype argument that accepts np.float32 or np.float64 to produce either single or double precision uniform random variables for select distributions. integers accepts a dtype argument with any signed or unsigned integer dtype.
- Uniforms (random and integers)
- Normals (standard_normal)
- Standard Gammas (standard_gamma)
- Standard Exponentials (standard_exponential)

In [6]: rng = np.random.default_rng()

In [7]: rng.random(3, dtype=np.float64)
Out[7]: array([0.94414369, 0.62474448, 0.80607091])

In [8]: rng.random(3, dtype=np.float32)
Out[8]: array([0.4812641 , 0.4058587 , 0.32835156], dtype=float32)

In [9]: rng.integers(0, 256, size=3, dtype=np.uint8)
Out[9]: array([ 44, 219, 139], dtype=uint8)

Optional out argument that allows existing arrays to be filled for select distributions
- Uniforms (random)
- Normals (standard_normal)
- Standard Gammas (standard_gamma)
- Standard Exponentials (standard_exponential)
This allows multithreading to fill large arrays in chunks using suitable BitGenerators in parallel.

In [10]: rng = np.random.default_rng()

In [11]: existing = np.zeros(4)

In [12]: rng.random(out=existing[:2])
Out[12]: array([0.0811171 , 0.65147786])

In [13]: print(existing)
[0.0811171  0.65147786 0.         0.        ]

Optional axis argument for methods like choice, permutation and shuffle that controls which axis an operation is performed over for multi-dimensional arrays.

In [14]: rng = np.random.default_rng()

In [15]: a = np.arange(12).reshape((3, 4))

In [16]: a
Out[16]: 
array([[ 0,  1,  2,  3],
       [ 4,  5,  6,  7],
       [ 8,  9, 10, 11]])

In [17]: rng.choice(a, axis=1, size=5)
Out[17]: 
array([[ 3,  1,  1,  2,  2],
       [ 7,  5,  5,  6,  6],
       [11,  9,  9, 10, 10]])

In [18]: rng.shuffle(a, axis=1)        # Shuffle in-place

In [19]: a
Out[19]: 
array([[ 1,  0,  3,  2],
       [ 5,  4,  7,  6],
       [ 9,  8, 11, 10]])

Added a method to sample from the complex normal distribution (complex_normal)