[highd]

You may be able to do that. That's sort of a matter of preference which you'd like to do. If the two datasets share more structure than it's more advantageous to share the network. There's also a bunch of hybrid approaches, i.e. pretraining on every city and then fine-tune each independently.