def create_model()

in models/model_factory.py [0:0]

• 34 lines of code
• 4 McCabe index (conditional complexity)

def create_model(args, arch=None):

    n_classes = int(getattr(args, 'n_classes', 1000))
    bn_momentum = getattr(args, 'bn_momentum', 0.1)
    bn_eps = getattr(args, 'bn_eps', 1e-5)

    dropout = getattr(args, 'dropout', 0)
    drop_connect = getattr(args, 'drop_connect', 0)

    if arch is None:
        arch = args.arch

    if arch == 'attentive_nas_dynamic_model':
        model = AttentiveNasDynamicModel(
            args.supernet_config,
            n_classes = n_classes, 
            bn_param = (bn_momentum, bn_eps),
        )
    elif arch == 'attentive_nas_static_model':
        supernet = AttentiveNasDynamicModel(
            args.supernet_config,
            n_classes = n_classes, 
            bn_param = (bn_momentum, bn_eps),
        )
        # load from pretrained models
        supernet.load_weights_from_pretrained_models(args.pareto_models.supernet_checkpoint_path)

        # subsample a static model with weights inherited from the supernet dynamic model
        supernet.set_active_subnet(
            resolution=args.active_subnet.resolution,
            width = args.active_subnet.width,
            depth = args.active_subnet.depth,
            kernel_size = args.active_subnet.kernel_size,
            expand_ratio = args.active_subnet.expand_ratio
        )
        model = supernet.get_active_subnet()

        # house-keeping stuff
        model.set_bn_param(momentum=bn_momentum, eps=bn_eps)
        del supernet
    else:
        raise ValueError(arch)

    return model