Diverse Policies Recovering via Pointwise Mutual Information Weighted Imitation Learning