Why learning rate is still considered as the hyper parameter?

Wow, thanks for providing more information.

So the callback is basically used to save the time, by efficiently decreasing the value of the LR based on the configuration provided in the constructor and training metadata from the fit method.

This is smart hack :smile: