I switched a long time ago when tensorflow decided to change a few of it's low level functions' parameter orders, for no f'ing reason. The only reply on the bug report was "why weren't you using named parameters". Became clear to me that the devs don't care about users.
As it should. I basically use it for a lot of things where I'd previously use NumPy. Unlike with TF, the cognitive load is very, very low, and you can actually debug things.
This is not accurate because the first breakout search is some Chevy dealer who happens to be named keras....