一、前言
本文旨在使用化合物分子的SMILES字符串進(jìn)行數(shù)據(jù)模型訓(xùn)練，對(duì)其水溶性的值進(jìn)行預(yù)測(cè)。
之前的文章《深度學(xué)習(xí) GNN圖神經(jīng)網(wǎng)絡(luò)（三）模型思想及文獻(xiàn)分類案例實(shí)戰(zhàn)》引用的Cora數(shù)據(jù)集只有一張圖，屬于圖神經(jīng)網(wǎng)絡(luò)的節(jié)點(diǎn)分類問(wèn)題。本文介紹的是多圖批量訓(xùn)練的線性回歸問(wèn)題，在文章最后也討論了圖分類問(wèn)題。
二、ESOL數(shù)據(jù)集
本文使用的是ESOL數(shù)據(jù)集，在文章《如何將化學(xué)分子SMILES字符串轉(zhuǎn)化為Pytorch圖數(shù)據(jù)結(jié)構(gòu)——ESOL分子水溶性數(shù)據(jù)集解析》中有詳細(xì)介紹，在此不作詳述。
三、加載數(shù)據(jù)集
from torch_geometric.datasets import MoleculeNet
dataset = MoleculeNet(root="data", name="ESOL")
print('num_features:',dataset.num_features)
print('num_classes:',dataset.num_classes)
print('num_node_features',dataset.num_node_features)
print("size:", len(dataset))
d=dataset[10]
print("Sample:", d)
print("Sample y:", d.y)
print("Sample num_nodes:",d.num_nodes)
print("Sample num_edges:",d.num_edges)
這里可以得到數(shù)據(jù)集的一些基本信息：
num_features: 9
num_classes: 734
num_node_features 9
size: 1128
Sample: Data(x=[6, 9], edge_index=[2, 12], edge_attr=[12, 3], smiles='O=C1CCCN1', y=[1, 1])
Sample y: tensor([[1.0700]])
Sample num_nodes: 6
Sample num_edges: 12
四、數(shù)據(jù)拆分
將數(shù)據(jù)集拆分為訓(xùn)練數(shù)據(jù)和測(cè)試數(shù)據(jù)：
from torch_geometric.loader import DataLoader
data_size = len(dataset)
batch_size = 128
train_data=dataset[:int(data_size*0.8)]
test_data=dataset[int(data_size*0.8):]
train_loader = DataLoader(train_data, batch_size=batch_size, shuffle=True)
test_loader = DataLoader(test_data, batch_size=len(test_data))
五、構(gòu)造模型
import torch
import torch.nn as nn
from torch_geometric.nn import GCNConv
import matplotlib.pyplot as plt
from torch_geometric.nn import global_mean_pool
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
hidden_channels = 64
class GNN(nn.Module):
def __init__(self):
# 初始化Pytorch父類
super().__init__()
self.conv1=GCNConv(dataset.num_node_features, hidden_channels)
self.conv2=GCNConv(hidden_channels, hidden_channels)
self.conv3 = GCNConv(hidden_channels, hidden_channels)
self.conv4 = GCNConv(hidden_channels, hidden_channels)
self.out = nn.Linear(hidden_channels, 1)
# 創(chuàng)建損失函數(shù)，使用均方誤差
self.loss_function = nn.MSELoss()
# 創(chuàng)建優(yōu)化器，使用Adam梯度下降
self.optimiser = torch.optim.Adam(self.parameters(), lr=0.005,weight_decay=5e-4)
# 訓(xùn)練次數(shù)計(jì)數(shù)器
self.counter = 0
# 訓(xùn)練過(guò)程中損失值記錄
self.progress = []
# 前向傳播函數(shù)
def forward(self, x, edge_index,batch):
x=x.to(device)
edge_index=edge_index.to(device)
batch=batch.to(device)
x=self.conv1(x, edge_index)
x=x.relu()
x=self.conv2(x, edge_index)
x=x.relu()
x=self.conv3(x, edge_index)
x=x.relu()
x=self.conv4(x, edge_index)
x=x.relu()
# 全局池化
x = global_mean_pool(x, batch) # [x, batch]
out=self.out(x)
return out
# 訓(xùn)練函數(shù)
def train(self, data):
# 前向傳播計(jì)算，獲得網(wǎng)絡(luò)輸出
outputs = self.forward(data.x.float(),data.edge_index,data.batch)
# 計(jì)算損失值
y=data.y.to(device)
loss = self.loss_function(outputs, y)
# 累加訓(xùn)練次數(shù)
self.counter += 1
# 每10次訓(xùn)練記錄損失值
if (self.counter % 10 == 0):
self.progress.append(loss.item())
# 每1000次輸出訓(xùn)練次數(shù)
if (self.counter % 1000 == 0):
print(f"counter={self.counter}, loss={loss.item()}")
# 梯度清零, 反向傳播, 更新權(quán)重
self.optimiser.zero_grad()
loss.backward()
self.optimiser.step()
# 測(cè)試函數(shù)
def test(self, data):
# 前向傳播計(jì)算，獲得網(wǎng)絡(luò)輸出
outputs = self.forward(data.x.float(),data.edge_index,data.batch)
# 把絕對(duì)值誤差小于1的視為正確，計(jì)算準(zhǔn)確度
y=data.y.to(device)
acc=sum(torch.abs(y-outputs)<1)/len(data.y)
return acc
# 繪制損失變化圖
def plot_progress(self):
plt.plot(range(len(self.progress)),self.progress)
六、訓(xùn)練模型
model = GNN()
model.to(device)
for i in range(1001):
for data in train_loader:
# print(data,'num_graphs:',data.num_graphs)
model.train(data)
counter=1000, loss=1.4304862022399902
counter=2000, loss=0.9842458963394165
counter=3000, loss=0.27240827679634094
counter=4000, loss=0.23295772075653076
counter=5000, loss=0.38499030470848083
counter=6000, loss=1.470423698425293
counter=7000, loss=0.845589816570282
counter=8000, loss=0.15707021951675415
繪制損失值變化圖：：
model.plot_progress()
七、測(cè)試結(jié)果
#torch.set_printoptions(precision=4,sci_mode=False) #pytorch不使用科學(xué)計(jì)數(shù)法顯示
for data in test_loader:
acc=model.test(data)
print(acc)
tensor([0.8186], device='cuda:0')
可以看到，預(yù)測(cè)值誤差小于1的占了81.86%，效果還行。
八、分類問(wèn)題
對(duì)于圖分類問(wèn)題，其實(shí)也差不多。只需要修改下Linear網(wǎng)絡(luò)層：
self.out = Linear(hidden_channels, dataset.num_classes)
這樣預(yù)測(cè)結(jié)果就會(huì)有num_classes個(gè)，取最大值的下標(biāo)索引即可。
偽代碼為：文章來(lái)源：http://www.zghlxwxcb.cn/news/detail-430799.html
pred=outputs.argmax(dim=1)
correct += int((pred == data.y).sum())
參考文獻(xiàn)
[1] https://pytorch-geometric.readthedocs.io/en/latest/get_started/colabs.html
[2] https://zhuanlan.zhihu.com/p/504978470文章來(lái)源地址http://www.zghlxwxcb.cn/news/detail-430799.html
到了這里，關(guān)于深度學(xué)習(xí) GNN圖神經(jīng)網(wǎng)絡(luò)（四）線性回歸之ESOL數(shù)據(jù)集水溶性預(yù)測(cè)的文章就介紹完了。如果您還想了解更多內(nèi)容，請(qǐng)?jiān)谟疑辖撬阉鱐OY模板網(wǎng)以前的文章或繼續(xù)瀏覽下面的相關(guān)文章，希望大家以后多多支持TOY模板網(wǎng)！
