Cityblock distance: Invalid arguments

[yannikschaelte]

Describe the bug

Using a ot.dist(a, b, metric="cityblock") is not possible, as in https://github.com/PythonOT/POT/blob/master/ot/utils.py#L224 an argument p is appended, which the underlying scipy implementation https://github.com/scipy/scipy/blob/v1.7.1/scipy/spatial/distance.py#L1798 does not appear to support. I believe that argument was only recently added.

To Reproduce

Steps to reproduce the behavior:

import scipy.spatial as spat
import ot

a = np.random.normal(size=(3, 1))
b = np.random.normal(size=(3, 1))

print(spat.distance.cdist(a, b, metric="cityblock"))
# fails
# print(spat.distance.cdist(a, b, metric="cityblock", p=2))

print(ot.dist(a, b, metric="sqeuclidean"))
# fails
# print(ot.dist(a, b, metric="cityblock"))

Error:

Traceback (most recent call last):
  File "/home/yannik/tmp/ot_err.py", line 13, in <module>
    print(ot.dist(a, b, metric="cityblock"))
  File "/home/yannik/pyenv/env/lib/python3.9/site-packages/ot/utils.py", line 224, in dist
    return cdist(x1, x2, metric=metric, p=p)
  File "/home/yannik/pyenv/env/lib/python3.9/site-packages/scipy/spatial/distance.py", line 2954, in cdist
    return cdist_fn(XA, XB, out=out, **kwargs)
TypeError: cdist_cityblock(): incompatible function arguments. The following argument types are supported:
    1. (x: object, y: object, w: object = None, out: object = None) -> numpy.ndarray

Invoked with: array([[ 0.60832356],
       [ 2.05290883],
       [-0.54136776]]), array([[ 1.04421186],
       [-1.4185131 ],
       [ 0.23999721]]); kwargs: out=None, p=2

Expected behavior

The wrapper around the corresponding scipy functions should be adapted to https://docs.scipy.org/doc/scipy/reference/generated/scipy.spatial.distance.cdist.html.

Environment (please complete the following information):

• OS (e.g. MacOS, Windows, Linux): Linux
• Python version: 3.9
• How was POT installed (source, pip, conda): pip, version 0.8.0, not occurring under 0.7.0
• scipy 1.7.1

[rflamary]

Good catch, surprisingly this bug does no appear in scipy 1.5 that I had on my machine I had to update it to get it.

We should do a PR to handle this case (sin p is used only for 'minkowski')