[MRG] Fix tests on numpy master

[lesteve]

numpy.apply_along_axis has changed behaviour when the function passed
in returns a 2d array. Very likely related to numpy/numpy#8441.

The first build in master I could find with the failure on numpy dev (from 2 days ago):
https://travis-ci.org/scikit-learn/scikit-learn/builds/200889357

import numpy as np
result = np.apply_along_axis(lambda X: np.array([[X[0]]]), 1, np.arange(6).reshape(3, -1))
print(result.shape)

outputs (3, 1, 1) on numpy master and (3, 1) on numpy 1.12.

A scikit-learn snippet closely related to the test failure:

from sklearn.gaussian_process import kernels
kernel = kernels.PairwiseKernel()
print(kernel.diag(np.arange(6).reshape(3, -1)).shape)

outputs (3, 1) on numpy master and (3,) on numpy 1.12.

[codecov]

Codecov Report

Merging #8355 into master will not change coverage.
The diff coverage is 100%.

@@           Coverage Diff           @@
##           master    #8355   +/-   ##
=======================================
  Coverage   94.75%   94.75%           
=======================================
  Files         342      342           
  Lines       60801    60801           
=======================================
  Hits        57609    57609           
  Misses       3192     3192

Impacted Files	Coverage Δ
sklearn/gaussian_process/kernels.py	91.54% <100%> (ø)	✅

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 05b5b37...3578d4b. Read the comment docs.

[lesteve]

The Travis build on this PR with numpy master is green as you can see from:
https://travis-ci.org/scikit-learn/scikit-learn/builds/201521405

[jnothman]

I think this is written in a confusing way, if self(arr) is returning an array of shape (1,1).

Shouldn't either of the following do the same?

np.apply_along_axis(lambda arr: self(arr)[0, 0], 1, X)
np.apply_along_axis(self, 1, X).ravel()

[lesteve]

I think this is written in a confusing way, if self(arr) is returning an array of shape (1,1).

Yeah I was not sure what was the best way to write it. Would you rather have

np.apply_along_axis(lambda arr: self(arr)[0, 0], 1, X)

Shouldn't either of the following do the same?

It does in numpy 1.12 but not in numpy master as I tried to highlight in the description.

[lesteve]

Actually my bad you need to use:the following in numpy master:

np.apply_along_axis(lambda arr: self(arr)[0], 1, X)

[jnothman]

Ravelling the output doesn't wok without the lambda?

…

On 15 February 2017 at 17:49, Loïc Estève ***@***.***> wrote: ***@***.**** commented on this pull request. ------------------------------ In sklearn/gaussian_process/kernels.py <#8355>: > @@ -1852,7 +1852,7 @@ def diag(self, X): Diagonal of kernel k(X, X) """ # We have to fall back to slow way of computing diagonal - return np.apply_along_axis(self, 1, X)[:, 0] + return np.apply_along_axis(lambda arr: self(arr).ravel(), 1, X)[:, 0] Actually my bad you need to use:the following in numpy master: np.apply_along_axis(lambda arr: self(arr)[0], 1, X) — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#8355>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAEz67UlTmgNOoezB-kMR_G3f53yW-0xks5rcp_dgaJpZM4MAfSB> .

[lesteve]

Ravelling the output doesn't wok without the lambda?

Ah sorry I read your comment too quickly above, this does work in numpy master indeed:

np.apply_along_axis(self, 1, X).ravel()

[jnothman]

LGTM

[lesteve]

Thanks!