[Numpy-discussion] Odd numerical difference between Numpy 1.5.1 and Numpy > 1.5.1

[Robert Kern]

On Mon, Apr 11, 2011 at 13:54, Skipper Seabold <jsseabold at gmail.com> wrote:
> All,
>
> We noticed some failing tests for statsmodels between numpy 1.5.1 and
> numpy >= 1.6.0. These are the versions where I noticed the change. It
> seems that when you divide a float array and multiply by a boolean
> array the answers are different (unless the others are also off by
> some floating point). Can anyone replicate the following using this
> script or point out what I'm missing?
>
> import numpy as np
> print np.__version__
> np.random.seed(12345)
>
> test = np.random.randint(0,2,size=10).astype(bool)
> t = 1.345
> arr = np.random.random(size=10)
>
> arr # okay
> t/arr # okay
> (1-test)*arr # okay
> (1-test)*t/arr # huh?

[~]
|12> 1-test
array([1, 0, 0, 0, 1, 0, 1, 1, 0, 1], dtype=int8)

[~]
|13> (1-test)*t
array([ 1.34472656,  0.        ,  0.        ,  0.        ,
1.34472656,  0.        ,
        1.34472656,  1.34472656,  0.        ,  1.34472656], dtype=float16)

Some of the recent ufunc changes or the introduction of the float16
type must have changed the way the dtype is chosen for the
"int8-array*float64-scalar" case. Previously, the result was a float64
array, as expected.

Mark, I expect this is a result of one of your changes. Can you take a
look at this?

-- 
Robert Kern

"I have come to believe that the whole world is an enigma, a harmless
enigma that is made terrible by our own mad attempt to interpret it as
though it had an underlying truth."
  -- Umberto Eco