Inconsistency between number of paramaters

[razvanc92]

I ran your implementation with a slightly different configuration (different dataset) and when using num_nodes: 87, enc_input_dim:1 your implementation has 223169 parameters while the original implementation with the same configuration has 371392. Do you have any ideas what could be the reason?

[xlwang233]

Hi, razvanc92
The reason that this inconsistency occurred is that I only implemented the "filter_type=laplacian" case, while the original implementation has considered all three different cases: "laplacian", "random_walk", "dual_random_walk". In fact, in the "laplacian" case, the length of self._support is 1, whereas in the "dual_random_walk" case, the length is 2.
Thank you for pointing this issue out. It actually took me a while to detect the reason. And I'll make an update to address this issue.

[xlwang233]

Hi, @razvanc92
I have modified the code. Now, when testing with the METR-LA dataset, the model produces exactly the same amount of parameters as the original implementation does.

[razvanc92]

@xlwang233 Thank you, I'll take a loot asap. Also there are some other things that could be improved/fixed. At the moment you can not run the code on cpu since there are some .cuda(), instead it could be .to(device). Also in the train.py the data loader dataset is hardcoded.