Re: [matplotlib-devel] hexbin and nans

[Andrew S. <str...@as...>]

John Hunter wrote:
> On Fri, Jul 18, 2008 at 2:49 PM, Andrew Straw <str...@as...> wrote:
>   
>> I just spent a little time getting hexbin to discard nans without
>> failing in mysterious ways.
>>
>> I'm sending 2 patches: one will detect nans and raise a suitable
>> exception. The other will automatically drop the nans and continue with
>> hexbin. The 2nd seems a little nicer in functionality, but the code is a
>> little more convoluted and hence more brittle, and so I'm torn as to
>> which I'd prefer. (The problem with just detecting nans from the
>> beginning is that it would cause an extra pass through the data in all
>> cases.)
>>     
>
> I think the nans should be dropped with no raise or warning, which is
> consistent with our handling elsewhere. I don't think your approach,
> which if I am reading this correctly assumes that is x or y has nans
> then the min or max will be nan, is correct.  Eg,
>
>   In [51]: x = np.random.rand(10)
>
>   In [52]: x[4] = np.nan
>
>   In [53]: x.min()
>   Out[53]: 0.37072898459621617
>
>   In [54]: x.max()
>   Out[54]: 0.79367039009185769
>
>
> I think min and max in the presence of nans is undefined.
True -- that slipped past me, since in my use case the nans were always
coming out for both min and max...
>   But you can
> just pay for the extra pass
>
> mask = np.isnan(x) & np.isnan(y)
> if mask.any():
>      # mask where
If you're happy with that extra cost, I'll modify
axes.delete_masked_points() so that hexbin and scatter are automatically
filtered in this way. Given the speed complaints we sometimes get, I was
hesitant to add another pass through the data.

-Andrew