Main reason for now is that the cosine similarity used as context function is not suitable for natural images like those in Imagenet. We are experimenting some extensions based on convolutions for such use cases.