Example: Transfer learning from pre-trained model

It would be useful to better understand how people might do transfer learning with deep learning frameworks and the rest of the SciPy/PyData stack. If this is cheap then this might not necessarily end up with any Dask involvement, which is OK. The pragmatism of the method probably merits investigation just to see if anything does come up. I suspect that a blogpost or binder example would be well appreciated by the community in general, if good and intuitive examples do not already exist.

This came out of conversation with @ogrisel