How to combine data batch with GCN?

🐛 Bug

Hi!, Thanks for your amazing framework. But recently, I am to using GCN to extract spatial-feature from node features in a batch way, but it got the error like this:

ValueError: `MessagePassing.propagate` only supports `torch.LongTensor` of shape 
`[2, num_messages]` or `torch_sparse.SparseTensor` for argument edge_index`.

The code I am using

The edge_index and edge_weight is defined as below:

graph_edges = torch.tensor(graph_edges, dtype=torch.long, device="cuda:0").t().contiguous()
graph_edges_b = torch.stack([graph_edges for x in range(location_t.shape[0])], dim=0)
print("graph_weight", graph_weight.shape)      # [16, 171]
print("graph_edges_b", graph_edges_b.shape) # [16, 2, 171]
print("feature_t", feature_t.shape)                     # [16, 19, 512] 

then

enc_feature = self.g_conv1(feature_t, graph_edges_b, graph_weight)

it feeds back the error as below:

ValueError: `MessagePassing.propagate` only supports `torch.LongTensor` of shape `[2, num_messages]` or `torch_sparse.SparseTensor` for argument `edge_index`.

For more details:

The batch_size is 16, and all the graph is fully-connected, but the weights is different for each graph.

[rusty1s]

You can concatenate features as follows to avoid the for-loop:

feature = torch.cat(feature_t, dim=0)
graph_weight = torch.cat(graph_weight, dim=0)
cumsum = 0
edge_indices = []
for edge_index, x in zip(graph_edges_b, feature_t):
    edge_indices.append(edge_index + cumsum)
    cumsum += x.size(0)
graph_edges = torch.cat(edge_indices, dim=-1)

hidden_enc_feature = F.relu(self.g_conv1(feature, graph_edges, graph_weight)
enc_feature = F.relu(self.g_conv2(hidden_enc_feature, graph_edges, graph_weight)
out_batch = enc_feature.view(batch_size, num_nodes, -1)

[TianhangWang]

I see! Thanks a lot! Hope you have a good day!