Development brief : Try to use multi-layer cnn to extract the overall appearance of the style reference picture, and use the VGG19 network directly here. In addition, three style losses are defined, and the weighted average of the three losses is minimized.