At present, param validation on Interval returns a TypeError when provided with a large integer (greater than 64 bits). This is because of the np.isnan call, which attempts to convert to a Numpy integer, of which no type exists greater than 64 bits. This issue came to light when trying to use numpy.SeedSequence().entropy (a 128-bit integer) as a random_state seed.

What does this implement/fix? Explain your changes.

A TypeError is raised unexpectedly:

>>> from sklearn.utils._param_validation import Interval
>>> from numbers import Integral
>>> interval = Interval(Integral, 0, 2, closed="neither")
>>> 2 ** 128 in interval
Traceback (most recent call last):
  File ".../.local/share/JetBrains/Toolbox/apps/PyCharm-C/ch-0/231.9011.38/plugins/python-ce/helpers/pydev/pydevconsole.py", line 364, in runcode
    coro = func()
  File "<input>", line 1, in <module>
  File ".../scikit-learn/sklearn/utils/_param_validation.py", line 481, in __contains__
    if not isinstance(val, Integral) and np.isnan(val):
TypeError: ufunc 'isnan' not supported for the input types, and the inputs could not be safely coerced to any supported types according to the casting rule ''safe''

Instead of calling np.isnan() on the value alone, it now first checks if the value is an integer (since integers cannot be nan). Tests have been added to check it works with large integers (including integers larger than those representible by a float, approx 2**1024 and larger).

After these changes, we have the following output, as expected:

>>> from sklearn.utils._param_validation import Interval
>>> from numbers import Integral
>>> interval = Interval(Integral, 0, 2, closed="neither")
>>> 2**128 in interval
False

Any other comments?

• If we don't want to support integers larger than those representible by a float, we can use math.isnan instead, which raises an OverflowError when conversion to float fails.
• The _NanConstraint will also throw an OverflowError if provided with a large integer (> 2**1024), but its only use is within MissingValues which already has an Integral class check