First of all, thank you for your work on this library. I just have noticed a few documentation inconsistencies that would be great if they could be fixed:
In the ASTGCN Class, it says that the edge_index needs to be provided on the initialization, but edge_index is not one of the inputs in the actual code.
In the DConv Class, the forward pass says that edge_weight is optional, but it is not optional in the code. It should be an OptTensor with a default setting equal to None.
In the METR-LA dataset, the target features do not return the time of day (as I would think would be preferred)
target.append((self.X[:, **0**, i + num_timesteps_in : j]).numpy())
But in the PEMS-BAY dataset, the time of day features are returned.
target.append((self.X[:, **:**, i + num_timesteps_in : j]).numpy())
This makes the two datasets inconsistent to work with, even though they cover the exact same task.
STGCN takes as input (Batch, Input Time Steps, Num Nodes, In Channels) even though most other networks take in (Batch, Num Nodes, In Channels, Input Time Steps). Could it be adjusted to be consistent with the other networks through a simple permutation operation?
First of all, thank you for your work on this library. I just have noticed a few documentation inconsistencies that would be great if they could be fixed:
target.append((self.X[:, **0**, i + num_timesteps_in : j]).numpy())
But in the PEMS-BAY dataset, the time of day features are returned.
target.append((self.X[:, **:**, i + num_timesteps_in : j]).numpy())
This makes the two datasets inconsistent to work with, even though they cover the exact same task.
Keep up the great work. Thanks.