Jax: Scalars passed into np.array should return 0-dim arrays

[alexbw]

import jax.numpy as np
import numpy as onp

onp.array(0).dtype # works, because onp.array(0) returns a 0-dim array
np.array(0).dtype # doesn't work, because np.array(0) returns an int

This showed up when trying to do lax.dynamic_update_slice(array, value, np.array(0))

[kroq-gar78]

124 says this was fixed, but I still get an error:

In [1]: import numpy as onp

In [2]: import jax.numpy as np

In [3]: onp.array(0).dtype
Out[3]: dtype('int64')

In [4]: np.array(0).dtype
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-4-05f7ef02f43f> in <module>
----> 1 np.array(0).dtype

AttributeError: 'int' object has no attribute 'dtype'

I'm on commit 6476d5ffc6bee0e8d9d7b54914565d798b8dda29.

[neerajprad]

I am not sure if this is a regression, but we noticed this recently as well when we found that scan does not work well with scalar arrays (https://github.com/pyro-ppl/numpyro/pull/91). The reason is that jax.numpy.array(1.) returns a python float instead of a DeviceArray unlike numpy.

cc. @mattjj in case this needs to be reopened.

[mattjj]

@kroq-gar78 that PR was closed but not merged, so this bug still isn't fixed and remains open.

@neerajprad we're going to check in a from-scratch rewrite of scan in the next week or two, though that new version will need testing and could have similar issues at first. Thanks for pointing that out!