Add ensemble for different architectures (#235)

        state = torch.load(filename, map_location=lambda s, l: default_restore_location(s, 'cpu'))

        state = _upgrade_state_dict(state)

        states.append(state)

    args = states[0]['args']

    if model_arg_overrides is not None:

        args = _override_model_args(args, model_arg_overrides)

    # build ensemble

    ensemble = []

    for state in states:

        args = state['args']

        if model_arg_overrides is not None:

            args = _override_model_args(args, model_arg_overrides)

        # build model for ensemble

        model = task.build_model(args)

        model.upgrade_state_dict(state['model'])

        model.load_state_dict(state['model'], strict=True)

        ensemble.append(model)

    return ensemble, args