Pre-Training Meta-Rule Selection Policy for Visual Generative Abductive Learning