GAN-based Intrinsic Exploration for Sample Efficient Reinforcement Learning