Set device fails when used after set backend

[stavros11]

Describe the bug
qibo.set_device() does not have the expected behavior when used in the same script with qibo.set_backend().

To Reproduce

import qibo
qibo.set_backend("custom")
qibo.set_device("/CPU:0")
print(qibo.get_device())

will print "/GPU:0" if executed in a machine with GPU enabled, as if the qibo.set_device instruction is ignored. If set_backend is not used then set_device works as intended.

[scarrazza]

I can reproduce this issue, and this sounds like the _CONSTRUCTED_BACKENDS and K are not in sync.

[scarrazza]

A simpler approach could be this: https://docs.python.org/3/faq/programming.html#id11

But probably will break many things, including user custom code.

[scarrazza]

Here some short version of the "singleton" approach we discussed today.

We place the following code in the __init__.py where my_backend_object is one of our backend inherited classes (e.g. TensorflowBackend):

class backend:
    my_backend_object = 0
    _instance = None

    def __new__(cls, *args, **kwargs):
        if cls._instance is None:
            cls._instance = super(backend, cls).__new__(
                cls, *args, **kwargs)
        return cls._instance

    def __repr__(self):
        return repr(self.my_backend_object)


K = backend()

Then we can import K as usual and get the right object. @stavros11 what do you think?

[stavros11]

Thanks for the details on this. I actually just pushed an update in #370 which I think is very close to this idea, excluding the __new__ method which I think can be easily added. Please have a look and if you agree I can continue with this approach.