pytorch geometric的GNN、GCN节点分类
# -*- coding: utf-8 -*-
import os
import torch
import torch.nn.functional as F
import torch.nn as nn
from torch_geometric.datasets import Planetoid
import torch_geometric.nn as pyg_nn
import torch_geometric.transforms as T
# load dataset
def get_data(folder="node_classify/cora", data_name="cora"):
# dataset = Planetoid(root=folder, name=data_name)
dataset = Planetoid(root=folder, name=data_name,
transform=T.NormalizeFeatures())
return dataset
# create the graph cnn model
class GraphCNN(nn.Module):
def __init__(self, in_c, hid_c, out_c):
super(GraphCNN, self).__init__()
self.conv1 = pyg_nn.GCNConv(in_channels=in_c, out_channels=hid_c)
self.conv2 = pyg_nn.GCNConv(in_channels=hid_c, out_channels=out_c)
def forward(self, data):
# data.x data.edge_index
x = data.x # [N, C]
edge_index = data.edge_index # [2 ,E]
hid = self.conv1(x=x, edge_index=edge_index) # [N, D]
hid = F.relu(hid)
out = self.conv2(x=hid, edge_index=edge_index) # [N, out_c]
out = F.log_softmax(out, dim=1) # [N, out_c]
return out
class OwnGCN(nn.Module):
def __init__(self, in_c, hid_c, out_c):
super(OwnGCN, self).__init__()
self.in_ = pyg_nn.SGConv(in_c, hid_c, K=2)
self.conv1 = pyg_nn.APPNP(K=2, alpha=0.1)
self.conv2 = pyg_nn.APPNP(K=2, alpha=0.1)
self.out_ = pyg_nn.SGConv(hid_c, out_c, K=2)
def forward(self, data):
x, edge_index = data.x, data.edge_index
x = self.in_(x, edge_index)
x = F.dropout(x, p=0.1, training=self.training)
x = F.relu(self.conv1(x, edge_index))
x = F.dropout(x, p=0.1, training=self.training)
x = F.relu(self.conv2(x, edge_index))
x = F.dropout(x, p=0.1, training=self.training)
x = self.out_(x, edge_index)
return F.log_softmax(x, dim=1)
# todo list
class YourOwnGCN(nn.Module):
pass
def analysis_data(dataset):
print("Basic Info: ", dataset[0])
print("# Nodes: ", dataset[0].num_nodes)
print("# Features: ", dataset[0].num_features)
print("# Edges: ", dataset[0].num_edges)
print("# Classes: ", dataset.num_classes)
print("# Train samples: ", dataset[0].train_mask.sum().item())
print("# Valid samples: ", dataset[0].val_mask.sum().item())
print("# Test samples: ", dataset[0].test_mask.sum().item())
print("Undirected: ", dataset[0].is_undirected())
def main():
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
cora_dataset = get_data()
# todo list
# my_net = GraphCNN(in_c=cora_dataset.num_features, hid_c=150, out_c=cora_dataset.num_classes)
my_net = OwnGCN(in_c=cora_dataset.num_features, hid_c=300, out_c=cora_dataset.num_classes)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
my_net = my_net.to(device)
data = cora_dataset[0].to(device)
optimizer = torch.optim.Adam(my_net.parameters(), lr=1e-2, weight_decay=1e-3)
"""
# model train
my_net.train()
for epoch in range(500):
optimizer.zero_grad()
output = my_net(data)
loss = F.nll_loss(output[data.train_mask], data.y[data.train_mask])
loss.backward()
optimizer.step()
_, prediction = output.max(dim=1)
valid_correct = prediction[data.val_mask].eq(data.y[data.val_mask]).sum().item()
valid_number = data.val_mask.sum().item()
valid_acc = valid_correct / valid_number
print("Epoch: {:03d}".format(epoch + 1), "Loss: {:.04f}".format(loss.item()),
"Valid Accuracy:: {:.4f}".format(valid_acc))
"""
# model test
my_net = torch.load("node_classify/best.pth")
my_net.eval()
_, prediction = my_net(data).max(dim=1)
target = data.y
test_correct = prediction[data.test_mask].eq(target[data.test_mask]).sum().item()
test_number = data.test_mask.sum().item()
train_correct = prediction[data.train_mask].eq(target[data.train_mask]).sum().item()
train_number = data.train_mask.sum().item()
print("==" * 20)
print("Accuracy of Train Samples: {:.04f}".format(train_correct / train_number))
print("Accuracy of Test Samples: {:.04f}".format(test_correct / test_number))
def test_main():
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
cora_dataset = get_data()
data = cora_dataset[0].to(device)
my_net = torch.load("node_classify/best.pth")
my_net.eval()
_, prediction = my_net(data).max(dim=1)
target = data.y
test_correct = prediction[data.test_mask].eq(target[data.test_mask]).sum().item()
test_number = data.test_mask.sum().item()
train_correct = prediction[data.train_mask].eq(target[data.train_mask]).sum().item()
train_number = data.train_mask.sum().item()
print("==" * 20)
print("Accuracy of Train Samples: {:.04f}".format(train_correct / train_number))
print("Accuracy of Test Samples: {:.04f}".format(test_correct / test_number))
if __name__ == '__main__':
test_main()
# main()
# dataset = get_data()
# analysis_data(dataset)
pytorch下GCN代码解读
def main():
print("hello world")
main()
import os.path as osp
import argparse
import torch
import torch.nn.functional as F
from torch_geometric.datasets import Planetoid
import torch_geometric.transforms as T
from torch_geometric.nn import GCNConv, ChebConv # noqa
#GCN用于提取图的特征参数然后用于分类
#数据集初始化部分
parser = argparse.ArgumentParser()
parser.add_argument('--use_gdc', action='store_true',
help='Use GDC preprocessing.')
args = parser.parse_args()#是否使用GDC优化
dataset = 'CiteSeer'#训练用的数据集
path = osp.join(osp.dirname(osp.realpath(__file__)), '..', 'data', dataset)#数据集存放位置
dataset = Planetoid(path, dataset, transform=T.NormalizeFeatures())#数据初始化类，其dataset的基类是一个torch.utils.data.Dataset对象
data = dataset[0]#只有一个图作为训练数据
#print(data)
#预处理和模型定义
if args.use_gdc:
gdc = T.GDC(self_loop_weight=1, normalization_in='sym',
normalization_out='col',
diffusion_kwargs=dict(method='ppr', alpha=0.05),
sparsification_kwargs=dict(method='topk', k=128,
dim=0), exact=True)
data = gdc(data)#图扩散卷积用于预处理
#搭建模型
class Net(torch.nn.Module):
#放置参数层(一般为可学习层，不可学习层也可放置，若不放置，则在forward中用functional实现)
def __init__(self):
super(Net, self).__init__()#在不覆盖Module的Init函数的情况下设置Net的init函数
self.conv1 = GCNConv(dataset.num_features, 16, cached=True,
normalize=not args.use_gdc)#第一层GCN卷积运算输入特征向量大小为num_features输出大小为16
#GCNConv的def init需要in_channel和out_channel(卷积核的数量)的参数，并对in_channel和out_channel调用linear函数，而该函数的作用为构建全连接层
self.conv2 = GCNConv(16, dataset.num_classes, cached=True,
normalize=not args.use_gdc)#第二层GCN卷积运算输入为16(第一层的输出)输出为num_class
# self.conv1 = ChebConv(data.num_features, 16, K=2)
# self.conv2 = ChebConv(16, data.num_features, K=2)
#实现模型的功能各个层之间的连接关系(具体实现)
def forward(self):
x, edge_index, edge_weight = data.x, data.edge_index, data.edge_attr#赋值data.x特征向量edge_index图的形状，edge_attr权重矩阵
x = F.relu(self.conv1(x, edge_index, edge_weight))#第一层用非线性激活函数relu
#x,edge_index,edge_weight特征矩阵，邻接矩阵，权重矩阵组成GCN核心公式
x = F.dropout(x, training=self.training)#用dropout函数防止过拟合
x = self.conv2(x, edge_index, edge_weight)#第二层输出
return F.log_softmax(x, dim=1)#log_softmax激活函数用于最后一层返回分类结果
#Z=log_softmax(A*RELU(A*X*W0)*W1)A连接关系X特征矩阵W参数矩阵
#得到Z后即可用于分类
#softmax(dim=1)行和为1再取log x为节点的embedding
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')#指定设备
model, data = Net().to(device), data.to(device)#copy model,data到device上
#优化算法
optimizer = torch.optim.Adam([
dict(params=model.conv1.parameters(), weight_decay=5e-4),#权重衰减避免过拟合
dict(params=model.conv2.parameters(), weight_decay=0)#需要优化的参数
], lr=0.01) # Only perform weight-decay on first convolution.
#lr步长因子或者是学习率
#模型训练
def train():
model.train()#设置成train模式
optimizer.zero_grad()#清空所有被优化的变量的梯度
F.nll_loss(model()[data.train_mask], data.y[data.train_mask]).backward()#损失函数训练参数用于节点分类
optimizer.step()#步长
@torch.no_grad()#不需要计算梯度，也不进行反向传播
#测试
def test():
model.eval()#设置成evaluation模式
logits, accs = model(), []
for _, mask in data('train_mask', 'val_mask', 'test_mask'):#mask矩阵，掩膜作用与之相应的部分不会被更新
pred = logits[mask].max(1)[1]#mask对应点的输出向量最大值并取序号1
acc = pred.eq(data.y[mask]).sum().item() / mask.sum().item()#判断是否相等计算准确度
accs.append(acc)
return accs
best_val_acc = test_acc = 0
#执行
for epoch in range(1, 201):
train()
train_acc, val_acc, tmp_test_acc = test()#训练准确率，实际输入的准确率，测试准确率
if val_acc > best_val_acc:
best_val_acc = val_acc
test_acc = tmp_test_acc
log = 'Epoch: {:03d}, Train: {:.4f}, Val: {:.4f}, Test: {:.4f}'#类型及保留位数
print(log.format(epoch, train_acc, best_val_acc, test_acc))#输出格式化函数'''
总结
以上为个人经验，希望能给大家一个参考，也希望大家多多支持编程网。
