triu generates nan when called on matrices with infinite values

[gaberosser]

In numpy 1.8.1, I've just found some strange behaviour.

Summary: when calling np.triu on a matrix, inf entries in the discarded region get converted to nan. This broke my code unexpectedly.

import numpy as np
a = np.zeros((4, 4))
a[1:, 2] = np.inf
a
>>>array([[  0.,   0.,   0.,   0.],
          [ inf,   0.,   0.,   0.],
          [ inf,   0.,   0.,   0.],
          [ inf,   0.,   0.,   0.]])

np.triu(a)
>>>array([[  0.,   0.,   0.,   0.],
          [ nan,   0.,   0.,   0.],
          [ nan,   0.,   0.,   0.],
          [ nan,   0.,   0.,   0.]])

Also posted on stackexchange: https://stackoverflow.com/questions/24678932/numpy-triu-generates-nan-when-called-on-matrices-with-infinite-values

[jaimefrio]

We could use boolean indexing instead of multiplication, and may even make it faster if done properly:

In [18]: arr = np.random.rand(1000, 1000)

In [19]: %timeit np.tri(1000) * arr # this is was np.triu currently does
100 loops, best of 3: 8.74 ms per loop

In [28]: %%timeit # this would be the boolean approach
   ....: mask = ~np.tri(1000, dtype=bool)
   ....: out = arr.copy()
   ....: out[mask] = 0
   ....:
100 loops, best of 3: 7 ms per loop

I have only done limited testing, but haven't found any faster way of doing this while keeping a pure Python implementation. Ideas are welcome! There may be an extra ms to shave if we change np.tri to take an inverse kwarg so that it produces the negated mask directly.

Anyway, if there is no major objection to this approach, I'll have a PR ready later today.