Learning to draw samples: with application to amortized maximum likelihood estimator for generative adversarial learning