np.random.binomial produces inconsistent output with same seed in 1.16 -> 1.17

Reproducing code example:

import numpy as np
np.random.seed(42)

np.random.binomial(1, np.linspace(0, 1, 1000), 1000).sum()

produces 515 on version 1.16.3 and 514 on version 1.17.0+

Error message:

Numpy/Python version information:

1.16.3 3.7.3 (default, Mar 26 2019, 21:43:19)
[GCC 8.2.1 20181127]

and

1.17.2 3.7.3 (default, Mar 26 2019, 21:43:19)
[GCC 8.2.1 20181127]

Simpler example:

In [44]: np.__version__
Out[44]: '1.16.4'
In [45]: np.random.seed(12345)
In [46]: np.random.binomial(1, [0, 0.5])
Out[46]: array([0, 0])

In [60]: np.__version__
Out[60]: '1.18.0.dev0+33d5f9c'
In [61]: np.random.seed(12345)
In [62]: np.random.binomial(1, [0, 0.5])
Out[62]: array([0, 1])

This appears to be the result of an explicit check for p = 0 in the new code:

numpy/numpy/random/src/distributions/distributions.c

Line 908 in e4de01d

if ((n == 0LL) || (p == 0.0f))

The old code did not have this shortcut, so it would make at least one more call to rk_double(state) than is made in the new code.

Looks like a regression, the call sequence of the old code should not have been changed unless the values resulted in a hang, in which case it probably wouldn't matter since no one would use those values.

@charris

@charris, I'm looking into fixing the regression.

@WarrenWeckesser

@WarrenWeckesser Great, I was just starting to investigate.

We do test the generated sequences for many things, but it seems that doesn't catch changes in corner cases. Good thing we no longer make that guarantee :)

BUG: random: Create a legacy implementation of random.binomial.

Thank you all for looking into this! I really appreciate the work you put into Numpy.

Navigation Menu

Repository navigation

np.random.binomial produces inconsistent output with same seed in 1.16 -> 1.17 #14522

Reproducing code example:

Error message:

Numpy/Python version information:

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

np.random.binomial produces inconsistent output with same seed in 1.16 -> 1.17 #14522

Description

Reproducing code example:

Error message:

Numpy/Python version information:

Activity

WarrenWeckesser commented on Sep 16, 2019

WarrenWeckesser commented on Sep 16, 2019

charris commented on Sep 16, 2019

WarrenWeckesser commented on Sep 16, 2019

charris commented on Sep 16, 2019

dburkhardt commented on Sep 23, 2019

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions