Comment by jph00

I'm intrigued!... The kernel calling overhead and lack of any GPU while/scan/map/etc for Pytorch seems like a limitation, but I guess on 2nd thoughts you can still do all the keras fit/predict stuff and auto-connecting up the layers.