[v2.0] Type Infer Error for MXNet NumPy Array in-place Operation

[barry-jin]

Description

As stated in array api specification, an in-place operation must not change the dtype or shape of the in-place array. But in MXNet NumPy Array library, some binary in-place operation will raise Type inconsistent error. For example:

>>> import mxnet as mx
>>> a1 = mx.np.array([1.0], dtype=onp.float32)
>>> b1 = mx.np.array([2.0], dtype=onp.float64)
>>> a1 *= b1
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/zhenghuj/GitHub/incubator-mxnet/python/mxnet/numpy/multiarray.py", line 268, in _wrap_mxnp_np_ufunc
    return func(x1, x2)
  File "/Users/zhenghuj/GitHub/incubator-mxnet/python/mxnet/numpy/multiarray.py", line 1104, in __imul__
    return multiply(self, other, out=self)
  File "/Users/zhenghuj/GitHub/incubator-mxnet/python/mxnet/util.py", line 645, in _wrap_np_binary_func
    return func(x1, x2, out=out)
  File "/Users/zhenghuj/GitHub/incubator-mxnet/python/mxnet/numpy/multiarray.py", line 3308, in multiply
    return _mx_nd_np.multiply(x1, x2, out)
  File "/Users/zhenghuj/GitHub/incubator-mxnet/python/mxnet/util.py", line 645, in _wrap_np_binary_func
    return func(x1, x2, out=out)
  File "/Users/zhenghuj/GitHub/incubator-mxnet/python/mxnet/ndarray/numpy/_op.py", line 1085, in multiply
    return _api_internal.multiply(x1, x2, out)
  File "mxnet/_ffi/_cython/./function.pxi", line 186, in mxnet._ffi._cy3.core.FunctionBase.__call__
  File "mxnet/_ffi/_cython/./function.pxi", line 120, in mxnet._ffi._cy3.core.FuncCall
  File "mxnet/_ffi/_cython/./function.pxi", line 108, in mxnet._ffi._cy3.core.FuncCall3
  File "mxnet/_ffi/_cython/./base.pxi", line 91, in mxnet._ffi._cy3.core.CALL
mxnet.base.MXNetError: MXNetError: Type inconsistent, Provided = float32, inferred type = float64

The inferred type for a1 is expected to be float32 because *= is in-place operation and we should not change the dtype of a1.

Official NumPy has the expected behavior.

>>> import numpy as onp
>>> a = onp.array([1.0], dtype=onp.float32)
>>> b = onp.array([2.0], dtype=onp.float64)
>>> a += b
>>> a
array([3.], dtype=float32)

To Solve This

Need to check if the operation is in-place or not in InferType function https://github.com/apache/incubator-mxnet/blob/8fd17cef2ee854239c6e66f2dd3b9467aa222f79/src/operator/numpy/np_elemwise_broadcast_op.cc#L47-L61

[barry-jin]

@NathanYyc FYI