Adversarial Learning of a Sampler Based on an Unnormalized Distribution