New test errors on Python 3 Travis bots

[arokem]

We're now seeing this across several PRs.

======================================================================
ERROR: dipy.reconst.tests.test_dti.test_mask
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/travis/build/nipy/dipy/venv/lib/python3.4/site-packages/nose/case.py", line 198, in runTest
    self.test(*self.arg)
  File "/home/travis/build/nipy/dipy/venv/lib/python3.4/site-packages/dipy/reconst/tests/test_dti.py", line 520, in test_mask
    assert_almost_equal(dtifit_w_mask.S0_hat[0, 0, 0], dtifit.S0_hat[0, 0, 0])
  File "/opt/python/3.4.6/lib/python3.4/unittest/case.py", line 841, in assertAlmostEqual
    if round(abs(second-first), places) == 0:
TypeError: type numpy.ndarray doesn't define __round__ method

Let's keep discussions about it here, so that we can all debug it together.

Is anyone already looking into this? @skoudoro ?

[arokem]

Googling for this, I'm a bit mystified. This seems to have been discussed a while back:

http://thread.gmane.org/gmane.comp.python.numeric.general/37179

But I imagine it must've been fixed since.

[aarya22]

Maybe use np.round instead of round?

https://stackoverflow.com/questions/41319292/error-typeerror-type-numpy-ndarray-doesnt-define-round-method

[arokem]

To answer @aarya22's question over on #1395, the issue is that we don't have control of this rounding operation, because it is happening inside of the assertion, that we are getting from numpy. So, we can't change the way round is called.

[aarya22]

Ah I see its in the unittest, not in our code.

[aarya22]

I came across this from the folks over at pyTorch, maybe something is being skipped in Travis?

pytorch/pytorch#4260

[skoudoro]

no clue for the moment, still investigating and it starts to make me crazy.

Trusty image change on Travis as you can see here and our error appear almost at the same time. I'm still not sure if there is a link between this 2 events.

[skoudoro]

Can you use the previous Trusty image on your last PR @arokem ?

[arokem]

Oh yeah - check it out. Here it only happens with the pre-release version:

https://travis-ci.org/nipy/dipy/jobs/323884050

And now, boom! So it definitely has to do with the 1.14 release of numpy (3 days ago...).

[arokem]

Take a look at #1400: maybe that fixes it?