support unsorted segment intervals

[bmcfee]

This is a bugfix for issue #202.

[bmcfee]

TODO:

• Support unsorted intervals in util.adjust_intervals
• In general, replace implicit start time intervals[0, 0] and end-time intervals[-1, 1] by calls to min() and max() unless we've explicitly sorted the data already.
• Explicit test for util.sort_intervals
• Rewrite util.interpolate_intervals to support discontinuities; this may be better done as a separate PR though.

[bmcfee]

I think we should punt on rewriting interpolate_intervals since it's non-essential to fixing the current issue with implicitly ordered timings. AFAIK, there's no need for label interpolation with incomplete interval coverage, so we can deal with that if/when it comes up.

Otherwise, I think this one's ready for CR from @craffel or @urinieto . (Should be an easy one.)

[bmcfee]

(PS: speedy resolution of this would be appreciated, as it's stalling a project i'm working on.)

[craffel]

This, and the max call below, could be made 2x faster as intervals[:, 0].min() and intervals[:, 1].max() respectively.

[bmcfee]

Nope 😁 the slice construction dominates runtime for the range of values we care about.

N = 100
x = np.random.randn(N, 2)

%timeit x.min()
100000 loops, best of 3: 1.9 µs per loop

%timeit x[:, 0].min()
100000 loops, best of 3: 1.92 µs per loop

[craffel]

Well, there you go.

[craffel]

Looks reasonable to me. I am not an expert on interpolate_intervals, so it might be useful if another user chimed in to make sure the change makes sense, but if it's doing what you want it to do it seems ok to me.

[bmcfee]

Cool, thanks.

I didn't make any substantive changes to interpolate_intervals (all previous tests still pass). The function itself is strange and a little broken IMO, but this isn't the right place to fix it.

The logic is: for each time point t, find the first interval k that starts after t, go back to the previous segment k-1, and apply that segment's label to t. This obviously requires ordered intervals to work, ie, interval[k-1] immediately precedes interval[k].

[craffel]

Ok, cool.

[bmcfee]

@craffel thanks for the CR. Unless you object, I think this is merge-ready.

[craffel]

Feel free to squash and merge.