Denoising-based Contractive Imitation Learning