Entropy-based Training Methods for Scalable Neural Implicit Sampler