GAT学习：PyG实现GAT（图注意力神经网络）网络（一）

本文主要是介绍GAT学习：PyG实现GAT（图注意力神经网络）网络（一），希望对大家解决编程问题提供一定的参考价值，需要的开发者们随着小编来一起学习吧！

PyG实现GAT网络

预备知识
代码分析
- 完整代码
- GAL层

注意！！！：本文的实现方法为笔者使用pyg的数据结构实现的，效果并不是最佳效果，pyg内部有封装好的GAT函数，使用pyg封装函数的方法请跳转下面，链接中文章的效果是可以达到论文效果的：
GAT学习：PyG实现GAT（使用PyG封装好的GATConv函数）（三）

目前PyG的教程几乎都是教怎么实现GCN的，但没找到GAT的PyG的实现，基本都是Pytorch实现。Paper需要，学习了GAT，为了保证和GCN用同一框架实现，所以用PyG实现了GAT，这里记录下来，用PyG搭建了GAT网络。

预备知识

1.GAT的原理移步这里向往的GAT，介绍的很详细。
2.PyG的基本操作移步这几篇：
GCN学习：Pytorch-Geometric教程（一）
GCN学习：Pytorch-Geometric教程（二）
GCN学习：用PyG实现自定义layers的GCN网络及训练（五）

代码分析

完整代码

import torch
import math
from torch_geometric.nn import MessagePassing
from torch_geometric.utils import add_self_loops,degree
from torch_geometric.datasets import Planetoid
import ssl
import torch.nn.functional as Fclass GAL(MessagePassing):def __init__(self,in_features,out_featrues):super(GAL,self).__init__(aggr='add')self.a = torch.nn.Parameter(torch.zeros(size=(2*out_featrues, 1)))torch.nn.init.xavier_uniform_(self.a.data, gain=1.414)  # 初始化# 定义leakyrelu激活函数self.leakyrelu = torch.nn.LeakyReLU()self.linear=torch.nn.Linear(in_features,out_featrues)def forward(self,x,edge_index):x=self.linear(x)N=x.size()[0]row，col=edge_indexa_input = torch.cat([x[row], x[col]], dim=1)print('a_input.size',a_input.size())# [N, N, 1] => [N, N] 图注意力的相关系数（未归一化）temp=torch.mm(a_input,self.a).squeeze()print('temp.size',temp.size())e = self.leakyrelu(temp)print('e',e)print('e.size', e.size())#e_all为同一个节点与其全部邻居的计算的分数的和，用于计算归一化softmaxe_all=torch.zeros(x.size()[0])count = 0for i in col:e_all[i]+=e[count]count=count+1print('e_all',e_all)for i in range(len(e)):e[i]=math.exp(e[i])/math.exp(e_all[col[i]])print('attention',e)print('attention.size',e.size())return self.propagate(edge_index,x=x,norm=e)def message(self, x_j, norm):print('x_j:', x_j)print('x_j.size', x_j.size())print('norm', norm)print('norm.size', norm.size())print('norm.view.size', norm.view(-1, 1).size())return norm.view(-1, 1) * x_jssl._create_default_https_context = ssl._create_unverified_context
dataset = Planetoid(root='Cora', name='Cora')
x=dataset[0].x
edge_index=dataset[0].edge_indexclass Net(torch.nn.Module):def __init__(self):super(Net, self).__init__()self.gal = GAL(dataset.num_node_features,16)def forward(self, data):x, edge_index = data.x, data.edge_indexx = F.dropout(x, training=self.training)x = self.gal(x, edge_index)print('x_gal',x.size())return F.log_softmax(x, dim=1)model=Net()
data=dataset[0]
out=Net()(data)
optimizer = torch.optim.Adam(model.parameters(), lr=0.01, weight_decay=5e-4)
model.train()
for epoch in range(1):optimizer.zero_grad()out = model(data)loss = F.nll_loss(out[data.train_mask], data.y[data.train_mask])loss.backward()optimizer.step()
model.eval()
_, pred = model(data).max(dim=1)
correct = int(pred[data.test_mask].eq(data.y[data.test_mask]).sum().item())
acc = correct/int(data.test_mask.sum())
print('Accuracy:{:.4f}'.format(acc))
>>>Accuracy:0.3880

GAL层

GAL层的写法思路参考GCN学习：用PyG实现自定义layers的GCN网络及训练（五）从而可以实现自定义网络层。所以核心内容还是编写init forward message函数。
在这里插入图片描述

GAL层要实现的工作：
1.进行特征映射
2.计算所有 $e$ _ij值
3.计算所有 $a l p h a$ _ij
4.加权求和

class GAL(MessagePassing):def __init__(in_features,out_featrues):#进行加权求和super(GAL,self).__init__(aggr='add')#定义attention参数aself.a = torch.nn.Parameter(torch.zeros(size=(2*out_featrues, 1)))torch.nn.init.xavier_uniform_(self.a.data, gain=1.414)  # 初始化# 定义leakyrelu激活函数self.leakyrelu = torch.nn.LeakyReLU()self.linear=torch.nn.Linear(in_features,out_featrues)def forward(self,x,edge_index):#特征映射x=self.linear(x)N=x.size()[0]col,row=edge_index#将相邻接点的特征拼接，然后计算e值a_input = torch.cat([x[row], x[col]], dim=1)print('a_input.size',a_input.size())#将规模压缩到一维temp=torch.mm(a_input,self.a).squeeze()print('temp.size',temp.size())e = self.leakyrelu(temp)print('e',e)print('e.size', e.size())#e_all为同一个节点与其全部邻居的计算的分数的和，用于计算归一化softmaxe_all=torch.zeros(x.size()[0])count = 0for i in col:e_all[i]+=e[count]count=count+1print('e_all',e_all)#计算alpha值for i in range(len(e)):e[i]=math.exp(e[i])/math.exp(e_all[col[i]])print('attention',e)print('attention.size',e.size())#传递信息return self.propagate(edge_index,x=x,norm=e)def message(self, x_j, norm):print('x_j:', x_j)print('x_j.size', x_j.size())print('norm', norm)print('norm.size', norm.size())print('norm.view.size', norm.view(-1, 1).size())#计算求和项return norm.view(-1, 1) * x_j
>>>a_input.size torch.Size([10556, 32])
temp.size torch.Size([10556])
e tensor([-0.0023, -0.0004, -0.0010,  ..., -0.0054, -0.0048, -0.0023],grad_fn=<LeakyReluBackward0>)
e.size torch.Size([10556])
e_all tensor([-0.0037,  0.7354,  0.1100,  ..., -0.0025,  0.0254, -0.0182],grad_fn=<CopySlices>)
attention tensor([1.0014, 1.0033, 1.0027,  ..., 1.0130, 1.0135, 1.0161],grad_fn=<CopySlices>)
attention.size torch.Size([10556])
x_j: tensor([[-0.0411,  0.0475, -0.0020,  ...,  0.1014,  0.1919,  0.0331],[-0.0411,  0.0475, -0.0020,  ...,  0.1014,  0.1919,  0.0331],[-0.0411,  0.0475, -0.0020,  ...,  0.1014,  0.1919,  0.0331],...,[-0.1486, -0.1743, -0.1428,  ...,  0.1968,  0.0718, -0.0176],[-0.1486, -0.1743, -0.1428,  ...,  0.1968,  0.0718, -0.0176],[-0.1486, -0.1743, -0.1428,  ...,  0.1968,  0.0718, -0.0176]],grad_fn=<IndexSelectBackward>)
x_j.size torch.Size([10556, 16])
norm tensor([1.0014, 1.0033, 1.0027,  ..., 1.0130, 1.0135, 1.0161],grad_fn=<CopySlices>)
norm.size torch.Size([10556])
norm.view.size torch.Size([10556, 1])
x_gal torch.Size([2708, 16])
a_input.size torch.Size([10556, 32])
temp.size torch.Size([10556])
e tensor([-0.0016, -0.0020, -0.0010,  ...,  0.2144,  0.0202, -0.0003],grad_fn=<LeakyReluBackward0>)
e.size torch.Size([10556])
e_all tensor([-0.0046,  0.1969,  0.4509,  ...,  0.1620, -0.0042,  0.3253],grad_fn=<CopySlices>)
attention tensor([1.0030, 1.0026, 1.0036,  ..., 0.8951, 0.7370, 0.7221],grad_fn=<CopySlices>)
attention.size torch.Size([10556])
x_j: tensor([[-0.1055, -0.0221,  0.0717,  ...,  0.0453,  0.0534,  0.0031],[-0.1055, -0.0221,  0.0717,  ...,  0.0453,  0.0534,  0.0031],[-0.1055, -0.0221,  0.0717,  ...,  0.0453,  0.0534,  0.0031],...,[ 0.0421,  0.0349, -0.0459,  ...,  0.1171,  0.0008,  0.0766],[ 0.0421,  0.0349, -0.0459,  ...,  0.1171,  0.0008,  0.0766],[ 0.0421,  0.0349, -0.0459,  ...,  0.1171,  0.0008,  0.0766]],grad_fn=<IndexSelectBackward>)
x_j.size torch.Size([10556, 16])
norm tensor([1.0030, 1.0026, 1.0036,  ..., 0.8951, 0.7370, 0.7221],grad_fn=<CopySlices>)
norm.size torch.Size([10556])
norm.view.size torch.Size([10556, 1])
x_gal torch.Size([2708, 16])
a_input.size torch.Size([10556, 32])
temp.size torch.Size([10556])
e tensor([ 0.2280,  0.2321, -0.0004,  ...,  0.1363,  0.3448,  0.0414],grad_fn=<LeakyReluBackward0>)
e.size torch.Size([10556])
e_all tensor([ 0.4597, -0.0024,  0.2359,  ...,  0.0669,  0.2952,  0.5938],grad_fn=<CopySlices>)
attention tensor([0.7932, 0.7964, 0.6312,  ..., 0.6329, 0.7796, 0.5756],grad_fn=<CopySlices>)
attention.size torch.Size([10556])
x_j: tensor([[-0.0510,  0.0875,  0.1096,  ..., -0.1464, -0.0774, -0.0326],[-0.0510,  0.0875,  0.1096,  ..., -0.1464, -0.0774, -0.0326],[-0.0510,  0.0875,  0.1096,  ..., -0.1464, -0.0774, -0.0326],...,[ 0.0554,  0.0655, -0.0448,  ..., -0.0251, -0.0492, -0.1602],[ 0.0554,  0.0655, -0.0448,  ..., -0.0251, -0.0492, -0.1602],[ 0.0554,  0.0655, -0.0448,  ..., -0.0251, -0.0492, -0.1602]],grad_fn=<IndexSelectBackward>)
x_j.size torch.Size([10556, 16])
norm tensor([0.7932, 0.7964, 0.6312,  ..., 0.6329, 0.7796, 0.5756],grad_fn=<CopySlices>)
norm.size torch.Size([10556])
norm.view.size torch.Size([10556, 1])