How can I get node embedding from the basic GCN encoder?

[songsong0425]

I think that this question is duplicated with this question (#4561), but I need more detail answer for the problem.
(Fortunately, I solve the previous problem by running the code in other server with ubuntu version 18.)

First, I prepared the PyG Data from networkx graph and got the dataset as below:

Data(edge_index=[2, 2475552], weight=[2475552], x=[1986, 1648])

The original networkx data was made from two csv files, one is adjacency matrix (row: node, column: node) and another is feature matrix (row: node, column: 1,648 of features).
Note that I don't have edge attributes in the graph.

Then, I ran the basic GCN using GCNConv in torch_geometric.nn.

import torch
import torch.nn.functional as F
from torch_geometric.nn import GCNConv

feature_dim = pyg_graph.num_node_features

class GCNEncoder(torch.nn.Module):
    def __init__(self):
        super(GCNEncoder, self).__init__()
        self.conv1 = GCNConv(feature_dim, 256)
        self.conv2 = GCNConv(256, 16)

    def forward(self, data):
        x, edge_index = data.x, data.edge_index

        x = self.conv1(x, edge_index)
        x = F.relu(x)
        x = self.conv2(x, edge_index)

        return x

Because I only want the node embeddings with 16 dimensions, I did not add other activation functions.

By using GCNEncoder, I tried run the model directly input the dataset, but I faced the error.

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") 
print("The model will be running on", device, "device\n") 

ppi_model = GCNEncoder().to(device)
ppi_data = ppi_pyg_graph.to(device)

ppi_out = ppi_model(ppi_data)

RuntimeError: "addmm_cuda" not implemented for 'Long'

As in #4561, I tried to add edge attributes, but I don't know where should I fix.
Can anyone suggest idea for this problem?
Thanks!

[rusty1s]

Do you have a minimal example to reproduce? What dtype does edge_index, edge_weight and x have? What does data.validate() report?

[songsong0425]

I hope this reply be helpful to you.
I run the below code to generate the PyG Data. If there are any errors, please notify to me.

import networkx as nx
import numpy as np
import pandas as pd
import torch
from torch_geometric.utils.convert import from_networkx
from torch_geometric.data import Data

ppi_feature=pd.read_csv('PPI_Features_Identity.csv')
ppi_edges=pd.read_csv('PPI_Adjacency.csv', index_col=0)

ppi_feature = ppi_feature.rename({'Unnamed: 0':'Gene'}, axis=1)

#PPI graph construction
ppi_G = nx.from_numpy_matrix(ppi_edges.values)
ppi_G = nx.relabel_nodes(ppi_G, dict(enumerate(ppi_edges.columns)))
ppi_G = nx.relabel_nodes(ppi_G, { n:str(n) for n in ppi_G.nodes()})

ppi_feature.Gene = ppi_feature.Gene.astype(str)
node_attr = ppi_feature.set_index('Gene').to_dict('index')
nx.set_node_attributes(ppi_G, node_attr)

ppi_pyg_graph = from_networkx(ppi_G, group_node_attrs=all)
print(ppi_pyg_graph)

# Data(edge_index=[2, 371078], weight=[371078], x=[1986, 1986])
# There was error when I ran the data.validate(); AttributeError: 'GlobalStorage' object has no attribute 'validate'

Then I ran the GCNEncoder code.
I'm sorry to bother you..

[rusty1s]

Looks like the issue is that your x is of dtype == torch.long, while GCNConv expects floating point features. Simply doing x = x.float() solves this issue.

[songsong0425]

It works well! Thank you so much for your help!!
Can I ask one more question? When I get the embedding as tensor like below, how can I match the value to the corresponding node?
Does it follow the order of nodes in the adjacency matrix?

tensor([[-0.0010,  0.0002,  0.0021,  ...,  0.0007, -0.0014,  0.0016], -> which node's embedding?
        [-0.0023,  0.0001,  0.0027,  ...,  0.0020, -0.0018,  0.0023], 
        ...,
        [-0.0318, -0.0364,  0.0301,  ...,  0.0439,  0.0134,  0.0388]], 
       device='cuda:0', grad_fn=<AddBackward0>)

[rusty1s]

Yes, the order remains unchanged.

[songsong0425]

Thanks a lot! I fully understand the process. Have a nice day!