How to use GPSConv (with GENConv) in DeeperGCN

[lfangyu09]

Hi,

Thank you so much for providing this helpful package. May I ask a few questions about GPSConv:

1. I want to use GPSConv (with GENConv) in DeeperGCN.
   I checked two codes, graph_gps.py and ogbn_proteins_deepgcn.py.
   Could it work for using GPSConv (with GENConv) in DeeperGCN:

class DeeperGCN(torch.nn.Module):
    def __init__(self, hidden_channels, num_layers):
        super().__init__()

        self.node_encoder = Linear(data.x.size(-1), hidden_channels)
        self.edge_encoder = Linear(data.edge_attr.size(-1), hidden_channels)

        self.layers = torch.nn.ModuleList()
        for i in range(1, num_layers + 1):
            conv = GPSConv(hidden_channels, 
                             GENConv(hidden_channels, hidden_channels, aggr='softmax',
                                                       t=1.0, learn_t=True, num_layers=2, norm='layer'), 
                              heads=4,
                             attn_type='multihead', attn_kwargs={'dropout': 0.5})
            norm = LayerNorm(hidden_channels, elementwise_affine=True)
            act = ReLU(inplace=True)

            layer = DeepGCNLayer(conv, norm, act, block='res+', dropout=0.1,
                                 ckpt_grad=i % 3)
            self.layers.append(layer)

        self.lin = Linear(hidden_channels, data.y.size(-1))

    ......

2. In GPSConv, graph_gps.py:
   Is batch necessary? Could I not use batch?
   Is x = global_add_pool(x, batch) used for global feature? Should I remove this if I want to do a node task?

3. Add Position Encoding and Structural Encoding:
   Should the input of AddLaplacianEigenvectorPE and AddRandomWalkPE be the position of nodes (.pos) or node features (.x)?
   How to add Structural Encoding?

Thank you so much.