np.random.uniform documentation incorrect

This is in reference to: http://docs.scipy.org/doc/numpy/reference/generated/numpy.random.uniform.html

Consider this code:

import numpy as np

norm_min = np.zeros(1)
norm_max = np.zeros(1) + 10

print "Input type: {0}.".format(type(norm_min))

norm_init = np.random.uniform(low=norm_min, high=norm_max)

print "Output type: {0}.".format(type(norm_init))

print norm_init.tolist()

The documentation says the return type from this method is an ndarray, however, if the input arrays are single-element, the result is a float. Either the method is incorrect in returning a float (I believe if you give it a single-element array, that is what it should be returned), or else the documentation should note this.

Yes, this can be fixed with:

norm_init = np.random.uniform(low=norm_min, high=norm_max, size=norm_min.size)

but that does not address the contradiction above.

[demitri]

This also happens with np.random.normal. I’d guess others then too, but haven’t done an exhaustive search.

[rkern]

I recommend converting each parameter to a numpy array and returning a scalar iff all of the converted parameters have shape (). We should not explicitly check for numpy arrays since someone may pass a list, as in my example above.