Shared Transformer trained jointly on 浅間下交差点周辺 and 川崎市立幸高校周辺 road networks

This is the fifth installment in a series that builds traffic-flow prediction models from OpenStreetMap data and the UXsim simulator. Earlier rounds established a model for the 浅間下交差点周辺 area, tried transfer learning to a different network around 川崎市立幸高校, and then prototyped a multi-network model where one Transformer is shared across networks. That earlier shared model, however, had been trained only on 浅間下交差点周辺 data and simply reused for the other network. The new step is to actually train the Shared Transformer on both networks' data together. The author kept graph, position and fully-connected layers per network and shared only the Transformer and LayerNorm, starting from the previously trained weights. Because the two networks differ in size — 122 versus 126 nodes — a concern was that adapting too much to one network could degrade the other. The experiment did not show that degradation, which the author reads as a hint that the shared model is picking up traffic-flow patterns common to both networks rather than just each network's individual quirks. The next planned step is Batch Inference, processing multiple road networks in a single batch.