I'm not 100% happy with the design choices in the KipoiModel.
On one hand I'm sure most model will need pre- and postprocessing, as we see for DeepBind.
But maybe passing functions and arguments for these cases is not a good idea. If a user has to write these functions anyway, the distance to redefining predict on a child implementation is very short, and we should incentivise that instead.
I'm not 100% happy with the design choices in the KipoiModel.
On one hand I'm sure most model will need pre- and postprocessing, as we see for DeepBind. But maybe passing functions and arguments for these cases is not a good idea. If a user has to write these functions anyway, the distance to redefining predict on a child implementation is very short, and we should incentivise that instead.