model.py
import torch.nn as nn
from torch_geometric.nn import GCNConv
import torch.nn.functional as F
class gcn_cls(nn.Module):
def __init__(self,in_dim,hid_dim,out_dim,dropout_size=0.5):
super(gcn_cls,self).__init__()
self.conv1 = GCNConv(in_dim,hid_dim)
self.conv2 = GCNConv(hid_dim,hid_dim)
self.fc = nn.Linear(hid_dim,out_dim)
self.relu = nn.ReLU()
self.dropout_size = dropout_size
def forward(self,x,edge_index):
x = self.conv1(x,edge_index)
x = F.dropout(x,p=self.dropout_size,training=self.training)
x = self.relu(x)
x = self.conv2(x,edge_index)
x = self.relu(x)
x = self.fc(x)
return x
main.py
import torch
import torch.nn as nn
from torch_geometric.datasets import Planetoid
from model import gcn_cls
import torch.optim as optim
dataset = Planetoid(root='./data/Cora', name='Cora')
print(dataset[0])
cora_data = dataset[0]
epochs = 50
lr = 1e-3
weight_decay = 5e-3
momentum = 0.5
hidden_dim = 128
output_dim = 7
net = gcn_cls(cora_data.x.shape[1],hidden_dim,output_dim)
optimizer = optim.AdamW(net.parameters(),lr=lr,weight_decay=weight_decay)
#optimizer = optim.SGD(net.parameters(),lr = lr,momentum=momentum)
criterion = nn.CrossEntropyLoss()
print("****************Begin Training****************")
net.train()
for epoch in range(epochs):
out = net(cora_data.x,cora_data.edge_index)
optimizer.zero_grad()
loss_train = criterion(out[cora_data.train_mask],cora_data.y[cora_data.train_mask])
loss_val = criterion(out[cora_data.val_mask],cora_data.y[cora_data.val_mask])
loss_train.backward()
print('epoch',epoch+1,'loss-train {:.2f}'.format(loss_train),'loss-val {:.2f}'.format(loss_val))
optimizer.step()
net.eval()
out = net(cora_data.x,cora_data.edge_index)
loss_test = criterion(out[cora_data.test_mask],cora_data.y[cora_data.test_mask])
_,pred = torch.max(out,dim=1)
pred_label = pred[cora_data.test_mask]
true_label = cora_data.y[cora_data.test_mask]
acc = sum(pred_label==true_label)/len(pred_label)
print("****************Begin Testing****************")
print('loss-test {:.2f}'.format(loss_test),'acc {:.2f}'.format(acc))
参数设置
epochs = 50
lr = 1e-3
weight_decay = 5e-3
momentum = 0.5
hidden_dim = 128
output_dim = 7
output_dim是输出维度，也就是有多少可能的类别。
注意事项
1.发现loss不下降：
建议改一改lr(学习率)，我做的时候开始用的SGD,学习率设的0.01发现loss不下降，改成0.1后好了很多。如果用AdamW,0.001(1e-3)基本就够用了