图注意力网络pyGAT论文代码注解

83 阅读 0 评论 55 点赞

我是靠谱客的博主个性画笔，最近开发中收集的这篇文章主要介绍图注意力网络pyGAT论文代码注解，觉得挺不错的，现在分享给大家，希望可以做个参考。

概述

论文地址：https://arxiv.org/abs/1710.10903

代码地址：https://github.com/Diego999/pyGAT

由于都使用引文网络cora，大部分的工作类似，在这里没有对代码进行详细注解
关于数据加载，矩阵正则化，one-hot编码转换部分可以参考
https://blog.csdn.net/qq_44015059/article/details/105968961

对于邻接矩阵的处理，图注意力模型GAT用注意力机制替代了GCN中固定的标准化操作。本质上，GAT只是将原本GCN的标准化函数替换为使用注意力权重的邻居节点特征聚合函数

train.py

from __future__ import division
from __future__ import print_function
import os
import glob
import time
import random
import argparse
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.autograd import Variable
from utils import load_data, accuracy
from models import GAT, SpGAT
# Training settings
parser = argparse.ArgumentParser()
parser.add_argument('--no-cuda', action='store_true', default=False, help='Disables CUDA training.')
parser.add_argument('--fastmode', action='store_true', default=False, help='Validate during training pass.')
parser.add_argument('--sparse', action='store_true', default=False, help='GAT with sparse version or not.')
parser.add_argument('--seed', type=int, default=72, help='Random seed.')
parser.add_argument('--epochs', type=int, default=10000, help='Number of epochs to train.')
parser.add_argument('--lr', type=float, default=0.005, help='Initial learning rate.')
parser.add_argument('--weight_decay', type=float, default=5e-4, help='Weight decay (L2 loss on parameters).')
parser.add_argument('--hidden', type=int, default=4, help='Number of hidden units.')
parser.add_argument('--nb_heads', type=int, default=8, help='Number of head attentions.')
parser.add_argument('--dropout', type=float, default=0.6, help='Dropout rate (1 - keep probability).')
parser.add_argument('--alpha', type=float, default=0.2, help='Alpha for the leaky_relu.')
parser.add_argument('--patience', type=int, default=100, help='Patience')
args = parser.parse_args()
args.cuda = not args.no_cuda and torch.cuda.is_available()
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if args.cuda:
torch.cuda.manual_seed(args.seed)
# Load data
adj, features, labels, idx_train, idx_val, idx_test = load_data()
# Model and optimizer
if args.sparse:
model = SpGAT(nfeat=features.shape[1],
nhid=args.hidden,
nclass=int(labels.max()) + 1,
dropout=args.dropout,
nheads=args.nb_heads,
alpha=args.alpha)
else:
model = GAT(nfeat=features.shape[1],
nhid=args.hidden,
nclass=int(labels.max()) + 1,
dropout=args.dropout,
nheads=args.nb_heads,
alpha=args.alpha)
optimizer = optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
if args.cuda:
model.cuda()
features = features.cuda()
adj = adj.cuda()
labels = labels.cuda()
idx_train = idx_train.cuda()
idx_val = idx_val.cuda()
idx_test = idx_test.cuda()
features, adj, labels = Variable(features), Variable(adj), Variable(labels)
def train(epoch):
t = time.time()
model.train()
optimizer.zero_grad()
output = model(features, adj)
loss_train = F.nll_loss(output[idx_train], labels[idx_train])
acc_train = accuracy(output[idx_train], labels[idx_train])
loss_train.backward()
optimizer.step()
if not args.fastmode:
# Evaluate validation set performance separately,
# deactivates dropout during validation run.
model.eval()
output = model(features, adj)
loss_val = F.nll_loss(output[idx_val], labels[idx_val])
acc_val = accuracy(output[idx_val], labels[idx_val])
print('Epoch: {:04d}'.format(epoch+1),
'loss_train: {:.4f}'.format(loss_train.data.item()),
'acc_train: {:.4f}'.format(acc_train.data.item()),
'loss_val: {:.4f}'.format(loss_val.data.item()),
'acc_val: {:.4f}'.format(acc_val.data.item()),
'time: {:.4f}s'.format(time.time() - t))
return loss_val.data.item()
def compute_test():
model.eval()
output = model(features, adj)
loss_test = F.nll_loss(output[idx_test], labels[idx_test])
acc_test = accuracy(output[idx_test], labels[idx_test])
print("Test set results:",
"loss= {:.4f}".format(loss_test.data[0]),
"accuracy= {:.4f}".format(acc_test.data[0]))
# Train model
t_total = time.time()
loss_values = []
bad_counter = 0
best = args.epochs + 1
best_epoch = 0
for epoch in range(args.epochs):
loss_values.append(train(epoch))
torch.save(model.state_dict(), '{}.pkl'.format(epoch))
if loss_values[-1] < best:
best = loss_values[-1]
best_epoch = epoch
bad_counter = 0
else:
bad_counter += 1
if bad_counter == args.patience:
break
files = glob.glob('*.pkl')
for file in files:
epoch_nb = int(file.split('.')[0])
if epoch_nb < best_epoch:
os.remove(file)
files = glob.glob('*.pkl')
for file in files:
epoch_nb = int(file.split('.')[0])
if epoch_nb > best_epoch:
os.remove(file)
print("Optimization Finished!")
print("Total time elapsed: {:.4f}s".format(time.time() - t_total))
# Restore best model
print('Loading {}th epoch'.format(best_epoch))
model.load_state_dict(torch.load('{}.pkl'.format(best_epoch)))
# Testing
compute_test()

models.py

import torch
import torch.nn as nn
import torch.nn.functional as F
from layers import GraphAttentionLayer, SpGraphAttentionLayer
class GAT(nn.Module):
def __init__(self, nfeat, nhid, nclass, dropout, alpha, nheads):
"""Dense version of GAT."""
super(GAT, self).__init__()
self.dropout = dropout
# print('nfeat:',nfeat)
# 引文网络每个结点的特征向量:描述论文的单词词汇表 1433
# print('nhid:',nhid)
# print('nclass:',nclass) # 7类论文
# print('dropout:',dropout)
# 0.6
# print('alpha:',alpha)
# 0.2
# print('nheads:',nheads)
# 8
self.attentions = [GraphAttentionLayer(nfeat, nhid, dropout=dropout, alpha=alpha, concat=True) for _ in range(nheads)]
# 搭建8层GAT
for i, attention in enumerate(self.attentions):
self.add_module('attention_{}'.format(i), attention)
# print('self.add_module:',self.add_module)
# 考虑周边8个点的注意力机制，特征维度1433->4
# 输出相应的分类结果，由于综合考虑了周边八个点的注意力机制
# 输入维度为4*8=32，最终目的的分类结果为7类
# nhid * nheads 32
# nclass
7
# dropout
0.6
# alpha
0.2
self.out_att = GraphAttentionLayer(nhid * nheads, nclass, dropout=dropout, alpha=alpha, concat=False)
def forward(self, x, adj):
print('adj.shape:',adj.shape)
x = F.dropout(x, self.dropout, training=self.training)
# 综合考虑周边八个点的维度拼接
x = torch.cat([att(x, adj) for att in self.attentions], dim=1)
x = F.dropout(x, self.dropout, training=self.training)
# ELU激活函数
x = F.elu(self.out_att(x, adj))
return F.log_softmax(x, dim=1)
# 系数
class SpGAT(nn.Module):
def __init__(self, nfeat, nhid, nclass, dropout, alpha, nheads):
"""Sparse version of GAT."""
super(SpGAT, self).__init__()
self.dropout = dropout
self.attentions = [SpGraphAttentionLayer(nfeat,
nhid,
dropout=dropout,
alpha=alpha,
concat=True) for _ in range(nheads)]
for i, attention in enumerate(self.attentions):
self.add_module('attention_{}'.format(i), attention)
self.out_att = SpGraphAttentionLayer(nhid * nheads,
nclass,
dropout=dropout,
alpha=alpha,
concat=False)
def forward(self, x, adj):
x = F.dropout(x, self.dropout, training=self.training)
x = torch.cat([att(x, adj) for att in self.attentions], dim=1)
x = F.dropout(x, self.dropout, training=self.training)
x = F.elu(self.out_att(x, adj))
return F.log_softmax(x, dim=1)

utils.py

import numpy as np
import scipy.sparse as sp
import torch
def encode_onehot(labels):
classes = set(labels)
classes_dict = {c: np.identity(len(classes))[i, :] for i, c in enumerate(classes)}
labels_onehot = np.array(list(map(classes_dict.get, labels)), dtype=np.int32)
return labels_onehot
def load_data(path="./data/cora/", dataset="cora"):
"""Load citation network dataset (cora only for now)"""
print('Loading {} dataset...'.format(dataset))
idx_features_labels = np.genfromtxt("{}{}.content".format(path, dataset), dtype=np.dtype(str))
features = sp.csr_matrix(idx_features_labels[:, 1:-1], dtype=np.float32)
labels = encode_onehot(idx_features_labels[:, -1])
# build graph
idx = np.array(idx_features_labels[:, 0], dtype=np.int32)
idx_map = {j: i for i, j in enumerate(idx)}
edges_unordered = np.genfromtxt("{}{}.cites".format(path, dataset), dtype=np.int32)
edges = np.array(list(map(idx_map.get, edges_unordered.flatten())), dtype=np.int32).reshape(edges_unordered.shape)
adj = sp.coo_matrix((np.ones(edges.shape[0]), (edges[:, 0], edges[:, 1])), shape=(labels.shape[0], labels.shape[0]), dtype=np.float32)
# build symmetric adjacency matrix
adj = adj + adj.T.multiply(adj.T > adj) - adj.multiply(adj.T > adj)
features = normalize_features(features)
adj = normalize_adj(adj + sp.eye(adj.shape[0]))
idx_train = range(140)
idx_val = range(200, 500)
idx_test = range(500, 1500)
adj = torch.FloatTensor(np.array(adj.todense()))
features = torch.FloatTensor(np.array(features.todense()))
labels = torch.LongTensor(np.where(labels)[1])
idx_train = torch.LongTensor(idx_train)
idx_val = torch.LongTensor(idx_val)
idx_test = torch.LongTensor(idx_test)
return adj, features, labels, idx_train, idx_val, idx_test
def normalize_adj(mx):
"""Row-normalize sparse matrix"""
rowsum = np.array(mx.sum(1))
r_inv_sqrt = np.power(rowsum, -0.5).flatten()
r_inv_sqrt[np.isinf(r_inv_sqrt)] = 0.
r_mat_inv_sqrt = sp.diags(r_inv_sqrt)
return mx.dot(r_mat_inv_sqrt).transpose().dot(r_mat_inv_sqrt)
def normalize_features(mx):
"""Row-normalize sparse matrix"""
rowsum = np.array(mx.sum(1))
r_inv = np.power(rowsum, -1).flatten()
r_inv[np.isinf(r_inv)] = 0.
r_mat_inv = sp.diags(r_inv)
mx = r_mat_inv.dot(mx)
return mx
def accuracy(output, labels):
preds = output.max(1)[1].type_as(labels)
correct = preds.eq(labels).double()
correct = correct.sum()
return correct / len(labels)

layers.py

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
class GraphAttentionLayer(nn.Module):
"""
Simple GAT layer, similar to https://arxiv.org/abs/1710.10903
"""
def __init__(self, in_features, out_features, dropout, alpha, concat=True):
super(GraphAttentionLayer, self).__init__()
# print('in_features:',in_features)
# 引文网络每个结点的特征向量:描述论文的单词词汇表 1433(8)
32
# print('out_features:',out_features)
# 输出结点的特征向量维度
4(8)
7
# print('dropout:',dropout)
# 0.6
# print('alpha:',alpha)
# 0.2
self.dropout = dropout
self.in_features = in_features
self.out_features = out_features
self.alpha = alpha
self.concat = concat
self.W = nn.Parameter(torch.zeros(size=(in_features, out_features)))
# nn.init.xavier_uniform_每一层网络保证输入和输出的方差相同，此处的初始化方法使用的是均匀分布
nn.init.xavier_uniform_(self.W.data, gain=1.414)
# print('self.W.shape:',self.W.shape)
# torch.Size([1433, 4])(8)
torch.Size([32, 7])
self.a = nn.Parameter(torch.zeros(size=(2*out_features, 1)))
# self.a公式所表达的信息是对于节点i来说节点j的重要性，而忽略图结构性的信息(这个公式模型允许图中所有节点间计算相互间的影响而不是局限于k阶邻居节点)
# 维度推到可参考https://blog.csdn.net/qq_44015059/article/details/105749213
# nn.init.xavier_uniform_每一层网络保证输入和输出的方差相同，此处的初始化方法使用的是均匀分布
nn.init.xavier_uniform_(self.a.data, gain=1.414)
# print('self.a.shape:',self.a.shape)
# torch.Size([8, 1])(8)
torch.Size([14, 1])
self.leakyrelu = nn.LeakyReLU(self.alpha)
def forward(self, input, adj):
h = torch.mm(input, self.W)
N = h.size()[0]
# 2708是cora引文网络的节点数
# print('h.shape:',h.shape)
# 2708,4(8)
2708,7
a_input = torch.cat([h.repeat(1, N).view(N * N, -1), h.repeat(N, 1)], dim=1).view(N, -1, 2 * self.out_features)
# print('a_input.shape:',a_input.shape)
# torch.Size([2708, 2708, 8])(8)
torch.Size([2708, 2708, 14])
e = self.leakyrelu(torch.matmul(a_input, self.a).squeeze(2))
# e的维度
(2708,2708,8)*(8,1) ->(2708,2708,1)
zero_vec = -9e15*torch.ones_like(e)
# torch.where 三个输入参数，第一个是判断条件，第二个是符合条件的设置值，第三个是不满足条件的设置值
attention = torch.where(adj > 0, e, zero_vec)
attention = F.softmax(attention, dim=1)
attention = F.dropout(attention, self.dropout, training=self.training)
h_prime = torch.matmul(attention, h)
if self.concat:
return F.elu(h_prime)
else:
return h_prime
# 默认情况下，__repr__() 会返回和调用者有关的 “类名+object at+内存地址”信息
def __repr__(self):
return self.__class__.__name__ + ' (' + str(self.in_features) + ' -> ' + str(self.out_features) + ')'
class SpecialSpmmFunction(torch.autograd.Function):
"""Special function for only sparse region backpropataion layer."""
@staticmethod
def forward(ctx, indices, values, shape, b):
assert indices.requires_grad == False
a = torch.sparse_coo_tensor(indices, values, shape)
ctx.save_for_backward(a, b)
ctx.N = shape[0]
return torch.matmul(a, b)
@staticmethod
def backward(ctx, grad_output):
a, b = ctx.saved_tensors
grad_values = grad_b = None
if ctx.needs_input_grad[1]:
grad_a_dense = grad_output.matmul(b.t())
edge_idx = a._indices()[0, :] * ctx.N + a._indices()[1, :]
grad_values = grad_a_dense.view(-1)[edge_idx]
if ctx.needs_input_grad[3]:
grad_b = a.t().matmul(grad_output)
return None, grad_values, None, grad_b
class SpecialSpmm(nn.Module):
def forward(self, indices, values, shape, b):
return SpecialSpmmFunction.apply(indices, values, shape, b)
class SpGraphAttentionLayer(nn.Module):
"""
Sparse version GAT layer, similar to https://arxiv.org/abs/1710.10903
"""
def __init__(self, in_features, out_features, dropout, alpha, concat=True):
super(SpGraphAttentionLayer, self).__init__()
self.in_features = in_features
self.out_features = out_features
self.alpha = alpha
self.concat = concat
self.W = nn.Parameter(torch.zeros(size=(in_features, out_features)))
nn.init.xavier_normal_(self.W.data, gain=1.414)
self.a = nn.Parameter(torch.zeros(size=(1, 2*out_features)))
nn.init.xavier_normal_(self.a.data, gain=1.414)
self.dropout = nn.Dropout(dropout)
self.leakyrelu = nn.LeakyReLU(self.alpha)
self.special_spmm = SpecialSpmm()
def forward(self, input, adj):
dv = 'cuda' if input.is_cuda else 'cpu'
N = input.size()[0]
edge = adj.nonzero().t()
h = torch.mm(input, self.W)
# h: N x out
assert not torch.isnan(h).any()
# Self-attention on the nodes - Shared attention mechanism
edge_h = torch.cat((h[edge[0, :], :], h[edge[1, :], :]), dim=1).t()
# edge: 2*D x E
edge_e = torch.exp(-self.leakyrelu(self.a.mm(edge_h).squeeze()))
assert not torch.isnan(edge_e).any()
# edge_e: E
e_rowsum = self.special_spmm(edge, edge_e, torch.Size([N, N]), torch.ones(size=(N,1), device=dv))
# e_rowsum: N x 1
edge_e = self.dropout(edge_e)
# edge_e: E
h_prime = self.special_spmm(edge, edge_e, torch.Size([N, N]), h)
assert not torch.isnan(h_prime).any()
# h_prime: N x out
h_prime = h_prime.div(e_rowsum)
# h_prime: N x out
assert not torch.isnan(h_prime).any()
if self.concat:
# if this layer is not last layer,
return F.elu(h_prime)
else:
# if this layer is last layer,
return h_prime
def __repr__(self):
return self.__class__.__name__ + ' (' + str(self.in_features) + ' -> ' + str(self.out_features) + ')'