Open sahilm1992 opened 4 years ago
It is just an engineering trick by creating each road segment a local map between the local defined neighbor ids to the global road ids. For example, given the current road segment 13 and its three neighbors with road segment id 10, 40, 50, assuming the maximum is five, then you can create a local map (0 <-> 10, 1 <-> 40, 2 <-> 50, 3 <-> -1, 4 <-> -1) where -1 is a dummy element.
Alternatively, you can adopt a more simple approach by considering each road segment are virtually connected to all other segments and mask all virtual neighbors when you do decoding. This is easier to implement.
Thanks for the clarification.
Actually, this thing is causing confusion in my mind. for road segment 13, map is (0 <-> 10, 1 <-> 40, 2 <-> 50, 3 <-> -1, 4 <-> -1)
Now for road segment lets say 22, if map is (0 <-> 40, 1 <-> 10, 2 <-> 5, 3 <-> 50, 4 <-> -1)
So the same dimension will be used for different road_segment neighbors ?
Doesn't that cause any issue in convergence?
Note that the map is a locally defined to each road segment and when decoding the output dimension is conditional on the current road segment, in other words, it does not make sense to talk about output dimension without specifying which road segment it is conditioning on.
Hi,![image](https://user-images.githubusercontent.com/5204669/88487055-b0222980-cf9f-11ea-85af-4ed307934f6b.png)
In the above image, it is mentioned that a vector of size max( Nr) will be used.
Let us say we got P(r_i+1) for node 5 and node 10 where r_i is 5 or 10.
How do you know which dimension is for which neighbor?