BUG: ESPILON with np.finfo(float).eps outputs a TypeError

[bribritos]

Describe the issue:

Hi,

I have recently created a new python environment using the 3.10.4 python version.
I have installed the latest public version (1.22.3) and when I am trying to catch the epsilon, I have an error that cannot allow me to use the command: numpy.finfo(float).eps or numpy.finfo(numpy.float32).eps

This can be very inconvenient because I hwill use some solutions of Scikit-Learn that calls this.

Sorry for the inconvenience. Best.

Reproduce the code example:

import numpy as np
eps = np.finfo(float).eps

Error message:

Traceback (most recent call last):**

  File "C:\Users\ME\2022_deepl\lib\site-packages\numpy\core\getlimits.py", line 459, in __new__
    dtype = numeric.dtype(dtype)

TypeError: 'numpy.dtype[bool_]' object is not callable

During handling of the above exception, another exception occurred:

Traceback (most recent call last):

  File "C:\Users\ME\AppData\Local\Temp\ipykernel_10948\2225752278.py", line 1, in <cell line: 1>
    np.finfo(float).eps

  File "C:\Users\ME\2022_deepl\lib\site-packages\numpy\core\getlimits.py", line 462, in __new__
    dtype = numeric.dtype(type(dtype))

TypeError: 'numpy.dtype[bool_]' object is not callable

NumPy/Python version information:

1.22.3 3.10.4 (tags/v3.10.4:9d38120, Mar 23 2022, 23:13:41) [MSC v.1929 64 bit (AMD64)]

[seberg]

This very much sounds like gh-21008, are you running a Python debugger or profiler there?

I am a bit surprised that you would see the error in a Python shell with only those two lines of code, can you try that? Right now, I would assume that this is in an IPython cell and other code (e.g. scikit-image) may have run before, and debugging may be enabled for some reason?

There is a 1.21.4 release plans with a fix included, right now we are waiting on cython unfortunately. Cython did fix this already though.
Maybe ping @thomasjpfan just so make sure that the sklearn people know that they probably also have to do a new release when the new Cython version becomes available to avoid this (presumably) issue with cython+tracing (debugging/profiling).

EDIT: To be clear, the NumPy release will fix only the use caused by using np.random. But e.g. pandas, or scikit-learn willl have their own cython code and they need to release a version with a fixed Cython as well.
In most cases where this happens, NumPy is just the bystander where the error shows up, its not the cause.

[seberg]

To be clear – assuming this is the issue – the only good work-around is to avoid "tracing", or avoiding Python 3.10 for another few weeks (and then installing all the latest bug-fix releases).
Otherwise, you would have to recompile the affected packages with additional compile flags, locally...

One way to check for tracing may be to do:

import sys

print(sys.gettrace())

If this does not return None, you tracing is enabled. I think you can disable the tracing with:

sys.settrace(None)

which may help, but I don't think a normal Python session should have it enabled to begin with (unless you are enabling debugging).

[bribritos]

Hi back,

You are right, I was using the IPython of the Spyder IDE. Then, I tested two solutions:
I run a Python Shell and I got this:

(2022_deepl) C:\Users\ME>python
Python 3.10.4 (tags/v3.10.4:9d38120, Mar 23 2022, 23:13:41) [MSC v.1929 64 bit (AMD64)] on win32
Type "help", "copyright", "credits" or "license" for more information.

import numpy as np
np.finfo(float).eps
2.220446049250313e-16
np.finfo(np.float32).eps
1.1920929e-07

It is working well ...
The second solution has been to test when I start up spyder by directly using the IPython:

Python 3.10.4 (tags/v3.10.4:9d38120, Mar 23 2022, 23:13:41) [MSC v.1929 64 bit (AMD64)]
Type "copyright", "credits" or "license" for more information.

IPython 7.33.0 -- An enhanced Interactive Python.
import numpy as np
np.finfo(float).eps
Out[2]: 2.220446049250313e-16

As we can see, we are all ok.

I decide then to run my script (from Spyder) and I have this:

Traceback (most recent call last):
File "C:\Users\ME\2022_deepl\lib\site-packages\numpy\core\getlimits.py", line 459, in new
dtype = numeric.dtype(dtype)
TypeError: 'numpy.dtype[bool_]' object is not callable
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "C:\Users\ME\2022_deepl\lib\site-packages\spyder_kernels\customize\spyderpdb.py", line 776, in run
super(SpyderPdb, self).run(cmd, globals, locals)
File "C:\Users\ME\AppData\Local\Programs\Python\Python310\lib\bdb.py", line 597, in run
exec(cmd, globals, locals)
File "C:\Users\ME\2022_deepl\lib\site-packages\imblearn_init_.py", line 52, in
from . import combine
File "C:\Users\ME\2022_deepl\lib\site-packages\imblearn\combine_init_.py", line 5, in
from .smote_enn import SMOTEENN
File "C:\Users\ME\2022_deepl\lib\site-packages\imblearn\combine_smote_enn.py", line 10, in
from ..base import BaseSampler
File "C:\Users\ME\2022_deepl\lib\site-packages\imblearn\base.py", line 15, in
from .utils import check_sampling_strategy, check_target_type
File "C:\Users\ME\2022_deepl\lib\site-packages\imblearn\utils_init.py", line 7, in
from ._validation import check_neighbors_object
File "C:\Users\ME\2022_deepl\lib\site-packages\imblearn\utils_validation.py", line 15, in
from sklearn.neighbors.base import KNeighborsMixin
File "C:\Users\ME\2022_deepl\lib\site-packages\sklearn\neighbors_init.py", line 17, in
from .nca import NeighborhoodComponentsAnalysis
File "C:\Users\ME\2022_deepl\lib\site-packages\sklearn\neighbors_nca.py", line 20, in
from ..decomposition import PCA
File "C:\Users\ME\2022_deepl\lib\site-packages\sklearn\decomposition_init.py", line 8, in
from .nmf import NMF, non_negative_factorization
File "C:\Users\ME\2022_deepl\lib\site-packages\sklearn\decomposition_nmf.py", line 24, in
EPSILON = np.finfo(np.float32).eps
File "C:\Users\ME\2022_deepl\lib\site-packages\numpy\core\getlimits.py", line 462, in new
dtype = numeric.dtype(type(dtype))
TypeError: 'numpy.dtype[bool]' object is not callable

Then, I run in the IPython of Spyder, and the issue comes back.

Traceback (most recent call last):
File "C:\Users\ME\2022_deepl\lib\site-packages\numpy\core\getlimits.py", line 459, in new
dtype = numeric.dtype(dtype)
TypeError: 'numpy.dtype[bool_]' object is not callable

During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "C:\Users\ME\AppData\Local\Temp\ipykernel_7544\2731468296.py", line 1, in <cell line: 1>
np.finfo(np.float32).eps
File "C:\Users\ME\2022_deepl\lib\site-packages\numpy\core\getlimits.py", line 462, in new
dtype = numeric.dtype(type(dtype))
TypeError: 'numpy.dtype[bool_]' object is not callable

I agree with you. I will forward this to the different packages (Spyder and Scikit-learn at least). Sorry for the inconvenience!

[seberg]

Ohhh, if Spyder installs tracing functions by default, than this may be a pretty pressing thing :/.

I don't think Spyder uses Cython, so they are probably fine, here. I pinged Thomas from sklearn, probably good enough :).
Once there is a new Cython release, I am sure a wave of bug-fix updates will follow quickly!

[seberg]

Thanks for the report and following up quickly! I will close this, since there is a duplicate and fixing it is already on the horizon.