search

benedekrozemberczki / pytorch_geometric_temporal

PyTorch Geometric Temporal: Spatiotemporal Signal Processing with Neural Machine Learning Models (CIKM 2021)
MIT License
2.62k stars 369 forks source link

GCN layers in TGCN #144

Closed jiayilong closed 2 years ago

jiayilong commented 2 years ago

Hi, I am wondering why we use different GCN layers to calculate X for the update gate, reset gate and candidate hidden state of GRU in the TGCN model?

In the source code, a GCN layer is created under each GRU gate, i.e., self.conv_z, self.conv_r and self.conv_h, so that there will be three different Xs for calculating these GRU gates and candidate hidden state.

I understand the X should be the same when calculating GRU gates and candidate hidden state. Can we calculate the X with a GCN layer first, and then feed the same X into a GRU to calculate gates?

Thank you.

benedekrozemberczki commented 2 years ago

Dear @jiayilong,

Yes you can do that. Can you star the repo?

Bests,

Benedek