🐛 Bug

The new complex number support introduces APIs that are incompatible with NumPy. Some of these functions existed in PyTorch 1.4, but they were practically useless and undocumented.

I'm concerned that this is going to make our life difficult going forward as we try to balance backwards compatibility with NumPy support. In general, we shouldn't be releasing new APIs that are incompatible with NumPy without good reasons.

Here are some examples I found with a few minutes of testing:

Tensor.real() is a function but ndarray.real is a property

>>> torch.randn(10, dtype=torch.complex64).real
<function Tensor.real>
>>> torch.randn(10, dtype=torch.complex64).numpy().real
array([ 0.1262403 , -0.40131333,  0.52708787, -0.10931669,  1.5780342 ,
       -0.17833038,  0.4660201 , -0.35434407,  0.73025763,  0.44405413],
      dtype=float32)

torch.real() (and Tensor.real()) returns a complex-type Tensor, but np.real (and ndarray.real) return floating-point type arrays:

>>> x = torch.randn(10, dtype=torch.complex64)
>>> torch.real(x).dtype
torch.complex64
>>> np.real(x.numpy()).dtype
dtype('float32')

cc @ezyang @gchanan @zou3519 @anjali411 @dylanbespalko