End-to-end Face-swapping via Adaptive Latent Representation Learning