Array initialization and indexing is inconsistent with official numpy

[Alicia1529]

Description

Two consistencies with official Numpy

1. when initializing a ndarray, the data type of official Numpy array will be casted to the input data type. DeepNumpy will set the default data type to be float32 unless the dtype parameter is explicitly declared.

>>> import numpy as onp
>>> from mxnet import numpy as np
>>> a = np.array([1, 2, 3, 4, 5])
>>> a
array([1., 2., 3., 4., 5.])
>>> a.dtype
dtype('float32')
>>> oa = onp.array([1, 2, 3, 4, 5])
>>> oa
array([1, 2, 3, 4, 5])
>>> oa.dtype
dtype('int64')

2. when indexing a 1-d array, the return type of numpy array is a value, while the return type of deenumpy array is still a deenumpy array.

a[0] # array([12.])
oa[0] # 12

[samskalicky]

@apeforest assign @mli

[sxjscience]

@Alicia1529 For the second question, I think you will need to use mx.npx.set_np() to enable numpy-compatible behavior.

import mxnet as mx
mx.npx.set_np()
a = np.array([1,2,3,4,5])
print(a[0]) # array(1.)

[yzhliu]

For the first first issue, I think it is intended as we try to make deep learning the first class. We will provide a global default dtype setting to switch the behavior.

[yzhliu]

default dtype landed in #18251