How do node features get passed around in MessagePassing base class?

[HongtaoYang]

I'm confused about how node features get passed around in MessagePassing base class? See the toy example bolow.

import numpy as np
import torch
from torch_geometric.nn import MessagePassing

# construct a simple graph, node features are 2 dimensional vectors whose value is simply the node id.
# e.g. for node 0, the feature is [0, 0]
edge_index = np.array([[0, 0, 0, 1, 1, 2, 2, 2, 2, 3, 4, 4, 5, 5], [1, 2, 3, 0, 2, 0, 1, 4, 5, 0, 2, 5, 2, 4]])
x = np.array([[0, 0], [1, 1], [2, 2], [3, 3], [4, 4], [5, 5]]) 

edge_index = torch.from_numpy(edge_index)
x = torch.from_numpy(x)

class MyMessagePassing(MessagePassing):
    def __init__(self, in_channels, out_channels):
        super(MyMessagePassing, self).__init__(aggr='add')
    
    def message(self, x_j: torch.Tensor) -> torch.Tensor:
        print(x)
        print(x_j)
        return x_j

    def forward(self, x, edge_index):
        x = x + 0.1  # toggle this does not change the print(x) output.
        self.propagate(edge_index, x=x)


gcn = MyMessagePassing(2, 2)
output = gcn(x, edge_index)

In the example, I simply inherit MessagePassing and add two print in the message method.

The problem is no matter what operations I apply on x during forward, the printed message is always the original input node features. I'm wondering where does the message method get the variable x from?

[rusty1s]

The x comes from global scope, and is therefore unmodified. Fixable by changing the message header to

def message(self, x: torch.Tensor, x_j: torch.Tensor):
    print(x)
    print(x_j)

[HongtaoYang]

Thanks!