draft: Joseba/multivariate quantile

[jdalch]

No description provided.

[M-Mouhcine]

Print traces should be remove. I suggest to replace with a logger from logging package.

[M-Mouhcine]

Great work ! Few remarks:

1. PEP8 recommends a line length of 79 characters. You should set your black formatter to that convention to avoid unneeded reformatting. Check this link.
2. Type hints should be consistent with older versions of python. Use Union[float, np.ndarray] instead or float or np.ndarray.
3. What is this condition axis % a.ndim == -1 supposed to check ?
4. Initializing axis to 0 or None seems to have no effect on the quantile computation, which is inconsistent with the docstring. Here is an example:

a = np.array([[1, 2, 3, 4, 5], [10, 20, 30, 40, 50], [100, 200, 300, 400, 500]])
q = np.array([0.1, 0.1, .5, .1, 0.9])
print(quantile_unweighted(a, q, axis=None)) # same result as quantile_unweighted(a, q, axis=0)

5. Following up on the last question, is there a case where axis > 0 ?

[jdalch]

Hello,

1. I can't find the option in viscose that your link says.
2. Ok.
3. As per the error message below, it checks that you are not trying to compute the quantile along the feature axis.
4. I changed the docstring. This is the expected behavior. For higher dimensional a, the behavior will be different if axis=None or if axis=0.
5. Yes, check line 534 of calibration.py.

[M-Mouhcine]

About point 3, if I understand correctly:

The condition axis % a.ndim == -1 is not correct to check if the user selected the feature axis. In Python, the modulo will always be positive because a.ndim is.

If you want to use modulo, you need to add a negative sign in front of a.ndim. This should work as intended: axis % -a.ndim == -1. However, I would rather pick a more expressive condition such as (a.ndim - axis == 1) or (axis == -1) or axis not in (a.ndim-1, -1).

[jdalch]

Ok, I corrected this.

[M-Mouhcine]

This should be removed.