np.float64(array) returns scalar for any single-element array (unlike np.float32, np.int64, etc.)

Hi all, I’m reporting a problem when using np.float64 as a constructor.

For any a = np.array([[scalar]]) assuming type(scalar) in (int, float), this raises AssertionError:

assert np.float64(a).shape is a.shape

The shape after casting to np.float64 is () instead of the expected (1,1). We can see what’s going on more clearly when only this last assertion fails:

scalar = 0.0 # example
a = np.array([[scalar]])
assert type(a) is np.ndarray
assert type(np.float64(a)) is np.ndarray #AssertionError

The problem is that np.float64(a) is np.float64(scalar). Contrary to expectations, np.float64(a) is not np.float64([[scalar]])

This happens with v1.14.2 as well as master. I’ve tested with python 3.6.4/2.7.11 on ubuntu 16.04/Mac OS 10.12.3.

[eric-wieser]

The difference is that unlike all the other floating types, issubclass(np.float64, float) is true, and it seems some of the float behavior leaks through.

[mhvk]

(Changed the title to reflect more clearly what I think the underlying issue is)