Comment by goolulusaurs

Do you think this could end up being implemented in TF and Pytorch?