Inconsistent type conversion from numpy.ndarray to mx.ndarray

[sxjscience]

When we create a mxnet ndarray using a numpy ndarray, the dtype will not be automatically inherited. However, if we create a mxnet ndarray based on an existing mxnet ndarray, the dtype will be inherited.

Example:

import mxnet as mx
import numpy as np
dat_np_to_mx = mx.nd.array(np.arange(10, dtype=np.int32))
dat_mx_to_mx =mx.nd.array(mx.nd.arange(10, dtype=np.int32)) 
print(dat_np_to_mx.dtype, dat_mx_to_mx.dtype)

Result:
<class 'numpy.float32'> <class 'numpy.int32'>

[ankkhedia]

@mxnet-label-bot : [Feature Request, Python]

[ankkhedia]

I tried reproducing the issue. Seems like whenever we want to try to create a ndarray from numpy array, the dtype defaults to float32 regardless of dtype of the input numpy array and this is according to the default definition here
https://github.com/apache/incubator-mxnet/blob/91730fdb6180ddcc3ab171df72c37f0ac15b59fd/python/mxnet/ndarray/ndarray.py#L2406
This is the reason for failure of this test too.
#12310

Will be good to have this feature.

[sxjscience]

This issue is solved by the new Numpy NDArray in MXNet:

import mxnet as mx
import numpy as np
mx.npx.set_np()
dat_np_to_mx = mx.np.array(np.arange(10, dtype=np.int32))
dat_mx_to_mx =mx.np.array(mx.np.arange(10, dtype=np.int32)) 
print(dat_np_to_mx.dtype, dat_mx_to_mx.dtype)

int32 int32