any reasons for inconsistent numpy arguments of numpy.zeros and numpy.random.randn

Question

I'm implementing a computation using numpy zeros and numpy.random.randn

W1 = np.random.randn(n_h, n_x) * .01
b1 = np.zeros((n_h, 1))

I'm not sure why random.randn() can accept two integers while zeros() needs a tuple. Is there a good reason for that?

Cheers, JChen.

Answer 1

Most likely it's just a matter of history. numpy results from the merger of several prior packages, and has a long development. Some quirks get cleaned up, others left as is.

randn(d0, d1, ..., dn)
zeros(shape, dtype=float, order='C')

randn has this note:

This is a convenience function. If you want an interface that takes a tuple as the first argument, use numpy.random.standard_normal instead.

standard_normal(size=None)

With * it is easy to pass a tuple to randn:

np.random.randn(*(1,2,3))

np.zeros takes a couple of keyword arguments. randn does not. You can define a Python function with a (*args, **kwargs) signature. But accepting a tuple, especially one with a common usage as shape, fits better. But that's a matter of opinion.

np.random.rand and np.random.random_sample are another such pair. Most likely rand and randn are the older versions, and standard_normal and random_sample are newer ones designed to conform to the more common tuple style.