Tensors inheriting from Tensor fail at math

[Jesse Moeller]

The overridden math methods defined in Tensor require methods which are only defined in DictTensor and DenseTensor. The Tensor base classes math operators are incomplete and inconsistent (case in point __add__ references Tensor.tensor_iadd, which raises AttributeError)

[Rob Speer (rspeer)]

Should they just raise a NotImplementedError instead?

Is there a reason you want to have an object that's neither a DictTensor nor a DenseTensor? I'm thinking that the *constructor* should raise a NotImplementedError because it's an abstract class.