Defdata does not produce terms of the correct type in some cases

[jfeser]

x = defop(torch.Tensor)
w = defop(float)
print(repr(x() * w()))
print(repr(w() * x()))

produces

_TensorTerm(_BaseOperation(<function _wrap_binop.<locals>._wrapped_op at 0x10658fd00>, name=mul, freshening=[]), (_TensorTerm(_BaseOperation(<function _.<locals>.func at 0x12a9dd7e0>, name=Tensor, freshening=[1515040756]), (), {}), _RealTerm(_BaseOperation(<function _.<locals>.func at 0x12a9dcf70>, name=float, freshening=[3341589068]), (), {})), {})
_RealTerm(_BaseOperation(<function _wrap_binop.<locals>._wrapped_op at 0x10658fd00>, name=mul, freshening=[]), (_RealTerm(_BaseOperation(<function _.<locals>.func at 0x12a9dcf70>, name=float, freshening=[3341589068]), (), {}), _TensorTerm(_BaseOperation(<function _.<locals>.func at 0x12a9dd7e0>, name=Tensor, freshening=[1515040756]), (), {})), {})

but both should be TensorTerms.

[jfeser]

This is a case where the simple type checking in typeof does not produce a correct result. mul is typed as (T * T) -> T and the simple type checking algorithm returns the type of the first argument.

[eb8680]

This is a case where the simple type checking in typeof does not produce a correct result. mul is typed as (T * T) -> T and the simple type checking algorithm returns the type of the first argument.

I think this is a problem with the resolution of * to a particular Operation, not with the typeof type inference logic. The __mul__ method of the concrete Term types should resolve to an Operation with the correct type for the two arguments - in particular, NumberTerm.__mul__(self, other) should return NotImplemented when not issubclass(typeof(other), numbers.Number) so that it can defer to TensorTerm.__rmul__, which will call the correct torch operation.

[jfeser]

I think this can be closed after #335

The example above still doesn't work, but it's due to a different issue. It's also not clear that we should try to replicate torch's behavior of casting floats to tensors.