Improving Image Autoencoder Embeddings with Perceptual Loss