How to deal with seq_len dimmension with GCNConv?

[Percibus]

I've just starting using PyG, and for my problem I have batched data with a time dimmension. My tensor has shape [batch, seq_len, nodes, features] . I want to apply the GCNConv layer separatly to each seq_len step.

I read on this issue

you need to permute your matrix to shape [batch_size, time_seq, num_nodes, num_feat]

But I'm not sure if I'm getting it correctly. Can I pass the tensor with this shape directly to the GCNConv layer and it will be applied to each seq_len step independenly, or should I reshape to [batch * nodes, seq_len * features] before feeding it into the layer?

Thanks in advance for the help!

[rusty1s]

Yes, you can pass this tensor directly, and both batch and seq_len dimensions are treated separately.

[Percibus]

Hi, thank you for your time! I assume is the same as reshaping to [batch * seq_len , nodes , features] as long as I use the default node_dim = -2 ?
And, when using operators that doesn't support batching directly, like GATv2Conv, I'm using:

x = x.view(batch_size * seq_len, nodes, features) # 48 nodes with 3 features each
data_list = [Data(x=x[i], edge_index=edge_index) for i in range(x.size(0))]
batch = Batch.from_data_list(data_list)  # and then pass batch.x and batch.edge_index to GAT      

Is this the correct approach to ensure each frame is processed as a distinct graph?
Thanks again for your answer.

[rusty1s]

Sorry for late reply. Yes, this is fully correct.

[Percibus]

Thanks!