Evaluating Parzen window log-likelihood for GAN

Question

I am trying to reimplement the original GAN paper by Ian Goodfellow et al. And I need to show that my implementation achieves same or similar results as the authors achieved