How to use RGCNConv and NNConv

[vlongle]

Hi,

I'm working on a grid-world problem with connected rooms. Each room is represented as a node and some doors between rooms might be blocked. As such, my data is

Data(edge_attr=[58, 2], edge_index=[2, 58], edge_type=[58], x=[24, 4])

where the edge_type indicates whether the edge is blocked and edge_attr has the following edge feature (blocked, number_of_doors). I tried to use RGCNConv and NNConv following pytorch_geometric/examples as follows

but neither of them worked. The RGCN returned this error

and NNConv this

Any help would be much appreciated!

[rusty1s]

Not really sure yet why the RGCNConv example does not work for you (can you create some custom data to reproduce?). In the second example, the nn needs to map to in_channels * out_channels of the NNConv:

nn1 = Sequential(Linear(num_edge_features, num_node_features * hidden_dim), ReLU()
self.conv1 = NNConv(num_node_features, hidden_dim, nn1, aggr='mean')
nn2 = Sequential(Linear(num_edge_features, hidden_dim * hidden_dim), ReLU()
self.conv2 = NNConv(hidden_dim, hidden_dim, nn1, aggr='mean')

[vlongle]

Here is a small example where both of my classes did not work

For the R-GCN, the error was

and for the NN-Conv (after I have added your fix)

Also, for the NNConv, I saw in the documentation that nn needs to output in_channels * out_channels but the example https://github.com/rusty1s/pytorch_geometric/blob/master/examples/mnist_nn_conv.py#L31 has nn1 outputs 32 which is also the out_channels in self.conv1 instead of having nn1 output d.num_features * 32 so that's why I was a bit confused.

Thanks!

[rusty1s]

I modified the examples in mnist_nn_conv.py to make that more clear. Thanks for pointing out.

Regarding your issues with RGCNConv: data.x should be of type float. With this, the following example works for me:

import networkx as nx
from torch_geometric.utils import from_networkx
from torch_geometric.nn import RGCNConv

G = nx.Graph()
G.add_edge(1, 2)
G.nodes[1]['x'] = [0., 0., 0.]
G.nodes[2]['x'] = [0., 1., 0.]

G.edges[(1, 2)]['edge_attr'] = [1., 0., 10., 5.]
G.edges[(1, 2)]['edge_type'] = 0

data = from_networkx(G)

conv = RGCNConv(in_channels=3, out_channels=6, num_relations=1)
out = conv(data.x, data.edge_index, data.edge_type)
print(out.shape)