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
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
|
.. _random-compatibility:
.. currentmodule:: numpy.random
Compatibility Policy
====================
`numpy.random` has a somewhat stricter compatibility policy than the rest of
NumPy. Users of pseudorandomness often have use cases for being able to
reproduce runs in fine detail given the same seed (so-called "stream
compatibility"), and so we try to balance those needs with the flexibility to
enhance our algorithms. :ref:`NEP 19 <NEP19>` describes the evolution of this
policy.
The main kind of compatibility that we enforce is stream-compatibility from run
to run under certain conditions. If you create a `Generator` with the same
`BitGenerator`, with the same seed, perform the same sequence of method calls
with the same arguments, on the same build of ``numpy``, in the same
environment, on the same machine, you should get the same stream of numbers.
Note that these conditions are very strict. There are a number of factors
outside of NumPy's control that limit our ability to guarantee much more than
this. For example, different CPUs implement floating point arithmetic
differently, and this can cause differences in certain edge cases that cascade
to the rest of the stream. `Generator.multivariate_normal`, for another
example, uses a matrix decomposition from ``numpy.linalg``. Even on the same
platform, a different build of ``numpy`` may use a different version of this
matrix decomposition algorithm from the LAPACK that it links to, causing
`Generator.multivariate_normal` to return completely different (but equally
valid!) results. We strive to prefer algorithms that are more resistant to
these effects, but this is always imperfect.
.. note::
Most of the `Generator` methods allow you to draw multiple values from
a distribution as arrays. The requested size of this array is a parameter,
for the purposes of the above policy. Calling ``rng.random()`` 5 times is
not *guaranteed* to give the same numbers as ``rng.random(5)``. We reserve
the ability to decide to use different algorithms for different-sized
blocks. In practice, this happens rarely.
Like the rest of NumPy, we generally maintain API source
compatibility from version to version. If we *must* make an API-breaking
change, then we will only do so with an appropriate deprecation period and
warnings, according to :ref:`general NumPy policy <NEP23>`.
Breaking stream-compatibility in order to introduce new features or
improve performance in `Generator` or `default_rng` will be *allowed* with
*caution*. Such changes will be considered features, and as such will be no
faster than the standard release cadence of features (i.e. on ``X.Y`` releases,
never ``X.Y.Z``). Slowness will not be considered a bug for this purpose.
Correctness bug fixes that break stream-compatibility can happen on bugfix
releases, per usual, but developers should consider if they can wait until the
next feature release. We encourage developers to strongly weight user’s pain
from the break in stream-compatibility against the improvements. One example
of a worthwhile improvement would be to change algorithms for a significant
increase in performance, for example, moving from the `Box-Muller transform
<https://en.wikipedia.org/wiki/Box%E2%80%93Muller_transform>`_ method of
Gaussian variate generation to the faster `Ziggurat algorithm
<https://en.wikipedia.org/wiki/Ziggurat_algorithm>`_. An example of
a discouraged improvement would be tweaking the Ziggurat tables just a little
bit for a small performance improvement.
.. note::
In particular, `default_rng` is allowed to change the default
`BitGenerator` that it uses (again, with *caution* and plenty of advance
warning).
In general, `BitGenerator` classes have stronger guarantees of
version-to-version stream compatibility. This allows them to be a firmer
building block for downstream users that need it. Their limited API surface
makes it easier for them to maintain this compatibility from version to version. See
the docstrings of each `BitGenerator` class for their individual compatibility
guarantees.
The legacy `RandomState` and the :ref:`associated convenience functions
<functions-in-numpy-random>` have a stricter version-to-version
compatibility guarantee. For reasons outlined in :ref:`NEP 19 <NEP19>`, we had made
stronger promises about their version-to-version stability early in NumPy's
development. There are still some limited use cases for this kind of
compatibility (like generating data for tests), so we maintain as much
compatibility as we can. There will be no more modifications to `RandomState`,
not even to fix correctness bugs. There are a few gray areas where we can make
minor fixes to keep `RandomState` working without segfaulting as NumPy's
internals change, and some docstring fixes. However, the previously-mentioned
caveats about the variability from machine to machine and build to build still
apply to `RandomState` just as much as it does to `Generator`.
|
