Did you train the decoder with cat5 settings, using only COCO dataset? And caculated feature Gram loss between reconstructed images and input images?
I'm suprised that the training process involve no style images, only COCO dataset as content images. And the Gram loss is not between stylized images and style images. Still, the network works well for style transfer
Did you train the decoder with cat5 settings, using only COCO dataset? And caculated feature Gram loss between reconstructed images and input images?
I'm suprised that the training process involve no style images, only COCO dataset as content images. And the Gram loss is not between stylized images and style images. Still, the network works well for style transfer