I think SciKit learn and Keras represent a decent level of abstraction for most real world use cases. You end up wanting a decent amount of control for most tas

I think SciKit learn and Keras represent a decent level of abstraction for most real world use cases. You end up wanting a decent amount of control for most tasks. More abstract would give you tools called things like AutoML and they do work, but you may have a hard time improving on what they do. They are an area of active research.