Hmmm looks like we still have failures on numpy-dev 😞:
https://travis-ci.org/lesteve/scikit-learn/builds/301861240

I think this PR should be merged anyway, since legacy=True is the right thing to do.

I compare the log in this PR and the log of the cron job yesterday, seems that this PR does not solve any failure(if we consider today's cron job as an accident?). Also, I don't have too much time to investigate but seems that this PR is related to your issue #8959?

I compare the log in this PR and the log of the cron job yesterday, seems that this PR does not solve any failure(if we consider today's cron job as an accident?)

Today's cron job is not an accident and is due to sign='legacy' -> legacy=True in the np.set_printoptions arguments. Basically as of now, the tests are not even run because of this.

With this PR, at least the tests are run. I agree it does not solve any failure and we will need to discuss with numpy to better understand how to deal with the failures.

Despite the failures in the Cron job I would be in favour of merging this PR.

Indeed, today's cron build is not an accident and this PR makes cron build able to run. Please kindly forgive me for the naive comment above.

No worries ! Naive comments are good sometimes, in this case it show that I probably did not explain myself well enough.

Going to merge this one.

@lesteve
Seems that #10074 is not fixed by the PR? Should we reopen it or do we need another issue for the 9 numpy-dev failures?

Seems that #10074 is not fixed by the PR? Should we reopen it or do we need another issue for the 9 numpy-dev failures?

Good catch I reopened it.

For anyone following, the argument was changed to legacy='1.13'

Thanks for following up. Yes we are using legacy='1.13' in our numpy-dev build (which passed this morning https://travis-ci.org/scikit-learn/scikit-learn/builds/305675956). We will be able to reenable doctests in our documentation when numpy/numpy#10059 is merged.