When passing a model_path to the FeatureExtractor, weights are loaded twice. One the pretrained once and then the given ones. Avoid the unnecessary load.