Description

The current mx.np.ndarray produces errors that cannot be tolerated in the face of simple division when the numerator and denominator data types have two differnt floating dtypes np.float32 and np.float64 such as

import mxnet as mx
import numpy as np
mx.npx.set_np()
numerator = mx.np.array(5.0).astype(np.float64)
denominator = mx.np.array(2.0)
numerator/denominator
>>>array(0., dtype=float64)

whereas the correct answer is obviously 2.5.

import numpy as np
numerator = np.array(5.0).astype(np.float64)
denominator = np.array(2.0)
numerator/denominator
>>>2.5

The same calculation steps are reliable under numpy, which forces us to face this problem directly.

Comments

@sxjscience