refactor(code): 使用 PyTorch 重构 BP 神经网络
- 将 BPNeuralNetwork 类从 NumPy 重新实现为 PyTorch 模型 - 使用 PyTorch 的自动求导和优化器替换手动反向传播和权重更新 - 将数据转换为 PyTorch 张量并支持 GPU 加速 -保留了原始代码的基本结构和功能
This commit is contained in:
parent
e00976c7dc
commit
99f9a797ba
@ -1,67 +1,30 @@
|
||||
import numpy as np
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.optim as optim
|
||||
from ucimlrepo import fetch_ucirepo
|
||||
from sklearn.model_selection import KFold
|
||||
from sklearn.metrics import accuracy_score
|
||||
from sklearn.preprocessing import StandardScaler
|
||||
|
||||
import numpy as np
|
||||
|
||||
# 定义BP神经网络类
|
||||
class BPNeuralNetwork:
|
||||
def __init__(self, input_size, hidden_size, output_size, learning_rate=0.1):
|
||||
class BPNeuralNetwork(nn.Module):
|
||||
def __init__(self, input_size, hidden_size, output_size):
|
||||
super(BPNeuralNetwork, self).__init__()
|
||||
# 初始化输入层、隐藏层和输出层的大小
|
||||
self.input_size = input_size
|
||||
self.hidden_size = hidden_size
|
||||
self.output_size = output_size
|
||||
self.learning_rate = learning_rate
|
||||
|
||||
# 初始化权重和偏置
|
||||
self.weights_input_hidden = np.random.randn(self.input_size, self.hidden_size) # 输入层到隐藏层的权重
|
||||
self.bias_hidden = np.random.randn(self.hidden_size) # 隐藏层的偏置
|
||||
self.weights_hidden_output = np.random.randn(self.hidden_size, self.output_size) # 隐藏层到输出层的权重
|
||||
self.bias_output = np.random.randn(self.output_size) # 输出层的偏置
|
||||
self.fc1 = torch.nn.Linear(self.input_size, self.hidden_size) # 输入层到隐藏层的全连接层
|
||||
self.fc2 = torch.nn.Linear(self.hidden_size, self.output_size) # 隐藏层到输出层的全连接层
|
||||
|
||||
def sigmoid(self, x):
|
||||
# Sigmoid激活函数
|
||||
return 1 / (1 + np.exp(-x))
|
||||
|
||||
def sigmoid_derivative(self, x):
|
||||
# Sigmoid激活函数的导数
|
||||
return x * (1 - x)
|
||||
|
||||
def forward(self, X):
|
||||
def forward(self, x):
|
||||
# 前向传播
|
||||
self.hidden_input = np.dot(X, self.weights_input_hidden) + self.bias_hidden # 隐藏层的输入
|
||||
self.hidden_output = self.sigmoid(self.hidden_input) # 隐藏层的输出
|
||||
self.final_input = np.dot(self.hidden_output, self.weights_hidden_output) + self.bias_output # 输出层的输入
|
||||
self.final_output = self.sigmoid(self.final_input) # 输出层的输出
|
||||
return self.final_output
|
||||
|
||||
def backward(self, X, y, output):
|
||||
# 反向传播
|
||||
output_error = y - output # 输出层的误差
|
||||
output_delta = output_error * self.sigmoid_derivative(output) # 输出层的误差信号
|
||||
|
||||
hidden_error = output_delta.dot(self.weights_hidden_output.T) # 隐藏层的误差
|
||||
hidden_delta = hidden_error * self.sigmoid_derivative(self.hidden_output) # 隐藏层的误差信号
|
||||
|
||||
# 更新权重和偏置
|
||||
self.weights_hidden_output += self.hidden_output.T.dot(output_delta) * self.learning_rate # 更新隐藏层到输出层的权重
|
||||
self.bias_output += np.sum(output_delta, axis=0) * self.learning_rate # 更新输出层的偏置
|
||||
self.weights_input_hidden += X.T.dot(hidden_delta) * self.learning_rate # 更新输入层到隐藏层的权重
|
||||
self.bias_hidden += np.sum(hidden_delta, axis=0) * self.learning_rate # 更新隐藏层的偏置
|
||||
|
||||
def train(self, X, y, epochs):
|
||||
# 训练网络
|
||||
for epoch in range(epochs):
|
||||
output = self.forward(X) # 前向传播
|
||||
self.backward(X, y, output) # 反向传播
|
||||
if epoch % 1000 == 0:
|
||||
loss = np.mean(np.square(y - output)) # 计算损失
|
||||
print(f'Epoch {epoch}, Loss: {loss}') # 打印损失
|
||||
|
||||
def predict(self, X):
|
||||
# 预测
|
||||
return np.round(self.forward(X)) # 四舍五入预测结果
|
||||
x = torch.sigmoid(self.fc1(x)) # 隐藏层的输出
|
||||
x = torch.sigmoid(self.fc2(x)) # 输出层的输出
|
||||
return x
|
||||
|
||||
# 将标签转换为one-hot编码
|
||||
def one_hot_encode(y):
|
||||
@ -72,7 +35,7 @@ def one_hot_encode(y):
|
||||
# 将标签转换为整数
|
||||
y_int = np.array([label_to_int[label] for label in y])
|
||||
n_values = np.max(y_int) + 1
|
||||
return np.eye(n_values)[y_int] # 返回one-hot编码
|
||||
return torch.eye(n_values)[y_int] # 返回one-hot编码
|
||||
|
||||
# fetch dataset
|
||||
wine_quality = fetch_ucirepo(id=186)
|
||||
@ -81,12 +44,6 @@ wine_quality = fetch_ucirepo(id=186)
|
||||
X = wine_quality.data.features.values # 特征数据
|
||||
y = wine_quality.data.targets.values # 标签数据
|
||||
|
||||
# metadata
|
||||
print(wine_quality.metadata)
|
||||
|
||||
# variable information
|
||||
print(wine_quality.variables)
|
||||
|
||||
# 特征缩放
|
||||
scaler = StandardScaler()
|
||||
X = scaler.fit_transform(X) # 标准化特征数据
|
||||
@ -94,6 +51,11 @@ X = scaler.fit_transform(X) # 标准化特征数据
|
||||
# 对标签进行one-hot编码
|
||||
y_encoded = one_hot_encode(y)
|
||||
|
||||
# 将数据转换为PyTorch张量并移至GPU
|
||||
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
||||
X = torch.tensor(X, dtype=torch.float32).to(device)
|
||||
y_encoded = y_encoded.to(device)
|
||||
|
||||
# 十折交叉验证
|
||||
kf = KFold(n_splits=10, shuffle=True, random_state=42)
|
||||
accuracies = []
|
||||
@ -103,12 +65,24 @@ for train_index, test_index in kf.split(X):
|
||||
y_train, y_test = y_encoded[train_index], y_encoded[test_index] # 训练集和测试集的标签数据
|
||||
|
||||
# 创建并训练BP神经网络
|
||||
nn = BPNeuralNetwork(input_size=X_train.shape[1], hidden_size=20, output_size=y_train.shape[1], learning_rate=0.0001) # 修改隐藏层大小和学习率
|
||||
nn.train(X_train, y_train, epochs=50000) # 增加训练轮数
|
||||
nn = BPNeuralNetwork(input_size=X_train.shape[1], hidden_size=20, output_size=y_train.shape[1]).to(device) # 修改隐藏层大小
|
||||
criterion = torch.nn.MSELoss() # 使用均方误差损失函数
|
||||
optimizer = optim.SGD(nn.parameters(), lr=0.0001) # 使用随机梯度下降优化器
|
||||
|
||||
for epoch in range(50000): # 增加训练轮数
|
||||
optimizer.zero_grad() # 清零梯度
|
||||
output = nn(X_train) # 前向传播
|
||||
loss = criterion(output, y_train) # 计算损失
|
||||
loss.backward() # 反向传播
|
||||
optimizer.step() # 更新权重和偏置
|
||||
|
||||
if epoch % 1000 == 0:
|
||||
print(f'Epoch {epoch}, Loss: {loss.item()}') # 打印损失
|
||||
|
||||
# 预测并计算准确率
|
||||
predictions = nn.predict(X_test)
|
||||
accuracy = accuracy_score(np.argmax(y_test, axis=1), np.argmax(predictions, axis=1)) # 计算准确率
|
||||
with torch.no_grad():
|
||||
predictions = nn(X_test)
|
||||
accuracy = accuracy_score(torch.argmax(y_test, dim=1).cpu().numpy(), torch.argmax(predictions, dim=1).cpu().numpy()) # 计算准确率
|
||||
accuracies.append(accuracy) # 存储每次交叉验证的准确率
|
||||
|
||||
print(f'Average Accuracy: {np.mean(accuracies)}') # 打印平均准确率
|
||||
print(f'Average Accuracy: {np.mean(accuracies)}') # 打印平均准确率
|
||||
Loading…
Reference in New Issue
Block a user