[Core] Fix type checking to be compatible with named tuples

[artbataev]

Signed-off-by: Vladimir Bataev [email protected]

What does this PR do ?

Fix inner type checking system for outputs to be compatible with named tuples.
This adds the ability to use named tuples (typing.NamedTuple or collections.namedtuple) for outputs. This is especially useful when the number of output elements is large.
Named tuples are compatible with simple tuples, there should be no problems with mixing them.

Usage

class OutputType(NamedTuple):
    y: torch.Tensor
    z: torch.Tensor

class MultipleOutputTypesWithNamedTuple(Typing):
    @property
    def output_types(self):
        return {"y": NeuralType(('B',), ElementType()), "z": NeuralType(('B',), ElementType())}

    @typecheck()
    def __call__(self, x):
        y = x + 1
        z = x + 2
        return OutputType(y=y, z=z)

obj = MultipleOutputTypesWithNamedTuple()
result = obj(x=torch.zeros(10))
result_y, result_z = result  # can unpack or use `result.y`, `result.z`

PR Type:

• New Feature
• Bugfix
• Documentation

Who can review?

@titu1994, @okuchaiev, @ericharper

[titu1994]

Fantastic ! Merge when ready

[ericharper]

LGTM. Thanks!