Multi-type Disentanglement without Adversarial Training