定义网络

  1. import torch
  2. from torch_geometric.nn import MessagePassing
  3. from torch_geometric.utils import add_self_loops, degree
  5. class GCNConv(MessagePassing):
  6.     def __init__(self, in_channels, out_channels):
  7.         super(GCNConv, self).__init__(aggr='add', flow='source_to_target')
  8.         # "Add" aggregation (Step 5).
  9.         # flow='source_to_target' 表示消息从源节点传播到目标节点
  10.         self.lin = torch.nn.Linear(in_channels, out_channels)
  12.     def forward(self, x, edge_index):
  13.         # x has shape [N, in_channels]
  14.         # edge_index has shape [2, E]
  16.         # Step 1: Add self-loops to the adjacency matrix.
  17.         edge_index, _ = add_self_loops(edge_index, num_nodes=x.size(0))
  19.         # Step 2: Linearly transform node feature matrix.
  20.         #将特征矩阵线性变化，本次的实验中变化：1433→16→7
  21.         print(f'线性变化之前的Xshape:{x.shape}')
  22.         #print(f'线性变化之前的X:{x}')
  23.         x = self.lin(x)
  24.         #print(f'变化后的x:{x}')
  25.         print(f'变化后的x.shape:{x.shape}')
  27.         # Step 3: Compute normalization.计算归一化
  28.         row, col = edge_index
  30.         #x.size(0)=2708，也就是结点的数量
  31.         #degree算出的是结点的度数，由于是对称矩阵，所以row和col的degree相等
  32.         deg = degree(col, x.size(0), dtype=x.dtype)
  33.         print(f'row:{row}')
  34.         print(f'col:{col}')
  35.         print(f'x.size(0):{x.size(0)}')
  36.         deg_inv_sqrt = deg.pow(-0.5)
  37.         print(f' deg_inv_sqrt:{ deg_inv_sqrt}')
  38.         norm = deg_inv_sqrt[row] * deg_inv_sqrt[col]
  39.         print(f'norm:{norm}')
  41.         # Step 4-5: Start propagating messages.
  42.         return self.propagate(edge_index, x=x, norm=norm)
  44.     def message(self, x_j, norm):
  45.         # x_j has shape [E, out_channels]
  46.         # Step 4: Normalize node features.
  47.         return norm.view(-1, 1) * x_j

初始化

数据集

  1. from torch_geometric.datasets import Planetoid
  2. import torch_geometric.transforms as T
  4. name_data = 'Cora'
  5. dataset = Planetoid(root='./data/', name=name_data)
  1. class Net(torch.nn.Module):
  2.     # torch.nn.Module 是所有神经网络单元的基类
  3.     def __init__(self):
  4.         super(Net, self).__init__()  ###复制并使用Net的父类的初始化方法，即先运行nn.Module的初始化函数
  5.         self.conv1 = GCNConv(dataset.num_node_features, 16)
  6.         self.conv2 = GCNConv(16, dataset.num_classes)
  8.     def forward(self, data):
  9.         x, edge_index = data.x, data.edge_index
  11.         x = self.conv1(x, edge_index)
  12.         x = F.relu(x)
  13.         x = F.dropout(x, training=self.training)
  14.         x = self.conv2(x, edge_index)
  16.         return F.log_softmax(x, dim=1)
  1. net=Net()
  2. out=net(dataset[0])

输出

由于定义了2个GCNConv层，所以输出可以看到变化的信息。
x.size(0)=2708，也就是结点的数量num_nodes
第一个：特征由1433→16，这是由于forward里的x = self.lin(x) （这里的self.lin = torch.nn.Linear(in_channels, out_channels)）
image.png
第二个：特征由16→7
image.png

rowshape:torch.Size([13264])
row和col的形状都是[13264]
deg。shape:torch.Size([2708])deg里面是结点的度

image.png