On Wed, Oct 2, 2013 at 11:43 AM, Charles R Harris <charlesr.harris@gmail.com> wrote:

Hi Stefan,

On Wed, Oct 2, 2013 at 9:29 AM, Stéfan van der Walt <stefan@sun.ac.za> wrote:

Hi Chuck

On Tue, Oct 1, 2013 at 1:07 AM, Charles R Harris
<charlesr.harris@gmail.com> wrote:
>> I'll bet the skimage problems come from
>> https://github.com/numpy/numpy/pull/3811. They may be doing something
>> naughty...
>>
>
> Reverting that commit fixes those skimage failures. However, there are a
> number of python2.7 failures that look pretty strange.

What is the exact change in behavior with that PR?  I'm trying to
figure out what skimage does wrong in this case.

The current master, and reverted for the 1.8 release only, is stricter about np.bool only taking values 0 or 1. Apparently the convolve returns boolean (I haven't checked) for boolean input, and consequently the check if the return value matches the number of 1 elements in the convolution kernel will fail when that number is greater than one. That is why the proposed fix is to view the boolean as uint8 instead. Note that out=(boolean) will still cause problems.

Chuck

So, just to be clear... what would happen if I had an array of floats between 0 and 1 inclusive and I cast that as a boolean using astype()?