BUG: np.random.multinomial(<float>, ...) raises a TypeError in numpy 1.26 #25061
Comments
|
Weird, on main I added an I would suspect it to be a cython 3 change about how it converts to integers which are not The |
# cython: language_level=3str
# distutils: define_macros=NPY_NO_DEPRECATED_API=NPY_1_7_API_VERSION
"""
Teasing out factors for numpy #25061, where passing a float to Cython code
[like np.random.RandomState.multinomial()] might raise
TypeError: 'float' object cannot be interpreted as an integer
depending whether this code is compiled with Cython==0.29 + numpy==1.25.2
vs. Cython==3.0 + numpy==1.26.1, which declare npy_intp differently.
"""
import numpy as np
cimport numpy as np
def func_int(int length):
"""Accepts an int or a float."""
return length
def func_Py_ssize_t(Py_ssize_t length):
"""Accepts an int; never a float."""
return length
def func_npy_intp(np.npy_intp length):
"""Accepts an int; float depends on Cython and numpy versions since they
define npy_intp differently."""
return length
# Test these in Python:
"""
import testcase as tc
tc.func_int(1.0) # should be OK
tc.func_Py_ssize_t(1.0) # should raise TypeError
tc.func_npy_intp(1.0) # depends on Cython 0.29 vs. 3.0
""" |
|
Thanks, so only Admittedly the definitions for intp seem outright wrong, not that it matters on most platforms... Unless somehow |
|
I think the story is:
==> So if __init__.cython-30.pxd reverted the definition of Whether that causes other problems, I can't say. A narrower fix would change multinomial to declare n as an |
|
That doesn't add up. NumPy has been the primary source for the definitions for a long time, unless |
|
Do ask the Cython folks for expertise on this and let me know where I was wrong. Here's a bit of info from a cython-users thread:
I'm unsure about when |
|
Maybe the difference is the language level default? |
|
I like a testable hypothesis! I wrote a unit test to test the limits. Results:
|
|
Another hypothesis test: @thalassemia verified that editing |
|
The |
|
The more thorough fix for this is: gh-25094 which fixes the underlying reason for why it changed, rather than a targeted specific one for multinomial. |
Describe the issue:
Up through numpy 1.25.2,
random.multinomial()accepted a float as the first argument n, whereas numpy 1.26 does not. (In our code, that value comes from a computation involvingscipy.constants.Avogadro.)In numpy 1.25.2 (tweaking an example from the docs):
In numpy 1.26.0 and 1.26.1:
Q. Is this an intentional API change?
I don't see it in the release notes.
n: int.np.npy_intp._ArrayLikeInt_co.npy_intpasPy_ssize_t.npy_intpasint.So I guess what changed was shifting API definitions from Cython's UFuncs.pyx to NumPy's __init__.cython-30.pxd
Environment: macOS 13.6, Intel CPU, Python 3.11.6.
Also: Ubuntu Linux, Intel CPU, Python 3.11.6.
Reproduce the code example:
Error message:
Runtime information:
import sys, numpy; print(numpy.__version__); print(sys.version)print(numpy.show_runtime())Context for the issue:
I'll adjust our calling code to convert these values to
int, so fixing this is not a priority.I'd like to:
The text was updated successfully, but these errors were encountered: