正则化方法(Regularization Techniques)
1. 目标
- 理解什么是过拟合及其影响
- 掌握常见正则化技术:L2 正则化、Dropout、Batch Normalization、Early Stopping
- 能够使用 PyTorch 编程实现这些正则化方法并进行比较分析
2. 数据构造与任务设定
本实验是一个带噪声的回归任务,目标函数为 y = x + N ( 0 , σ 2 ) y = x + \mathcal{N}(0, \sigma^2) y=x+N(0,σ2)。使用均匀分布采样输入 x ∈ [ − 1 , 1 ] x \in [-1, 1] x∈[−1,1]。
python
import numpy as np
import torch
import torch.utils.data as Data
N_SAMPLES = 20
NOISE_RATE = 0.4
train_x = np.linspace(-1, 1, N_SAMPLES)[:, np.newaxis]
train_y = train_x + np.random.normal(0, NOISE_RATE, train_x.shape)
validate_x = np.linspace(-1, 1, N_SAMPLES // 2)[:, np.newaxis]
validate_y = validate_x + np.random.normal(0, NOISE_RATE, validate_x.shape)
test_x = np.linspace(-1, 1, N_SAMPLES)[:, np.newaxis]
test_y = test_x + np.random.normal(0, NOISE_RATE, test_x.shape)
# 转换为 Tensor
train_x = torch.tensor(train_x, dtype=torch.float32)
train_y = torch.tensor(train_y, dtype=torch.float32)
validate_x = torch.tensor(validate_x, dtype=torch.float32)
validate_y = torch.tensor(validate_y, dtype=torch.float32)
test_x = torch.tensor(test_x, dtype=torch.float32)
test_y = torch.tensor(test_y, dtype=torch.float32)
train_dataset = Data.TensorDataset(train_x, train_y)
train_loader = Data.DataLoader(dataset=train_dataset, batch_size=10, shuffle=True)
3. 模型定义
3.1 原始 MLP(无正则化)
python
import torch.nn as nn
import torch.nn.init as init
class FC_Classifier(nn.Module):
def __init__(self, input_dim=1, hidden_dim=100, output_dim=1):
super().__init__()
self.fc1 = nn.Linear(input_dim, hidden_dim)
self.fc2 = nn.Linear(hidden_dim, output_dim)
self.activation = nn.ReLU()
self._init_weights()
def _init_weights(self):
init.normal_(self.fc1.weight, mean=0.0, std=0.1)
init.constant_(self.fc1.bias, 0)
init.normal_(self.fc2.weight, mean=0.0, std=0.1)
init.constant_(self.fc2.bias, 0)
def forward(self, x):
x = self.activation(self.fc1(x))
return self.fc2(x)
3.2 Dropout MLP
python
class DropoutMLP(nn.Module):
def __init__(self, dropout_rate=0.5):
super().__init__()
self.fc1 = nn.Linear(1, 100)
self.dropout = nn.Dropout(dropout_rate)
self.fc2 = nn.Linear(100, 1)
self.activation = nn.ReLU()
self._init_weights()
def _init_weights(self):
init.normal_(self.fc1.weight, mean=0.0, std=0.1)
init.constant_(self.fc1.bias, 0)
init.normal_(self.fc2.weight, mean=0.0, std=0.1)
init.constant_(self.fc2.bias, 0)
def forward(self, x):
x = self.dropout(self.fc1(x))
x = self.activation(x)
return self.fc2(x)
3.3 Batch Normalization MLP
python
class BNMLP(nn.Module):
def __init__(self):
super().__init__()
self.bn_input = nn.BatchNorm1d(1)
self.fc1 = nn.Linear(1, 100)
self.bn_hidden = nn.BatchNorm1d(100)
self.fc2 = nn.Linear(100, 1)
self.activation = nn.ReLU()
def forward(self, x):
x = self.bn_input(x)
x = self.fc1(x)
x = self.bn_hidden(x)
x = self.activation(x)
return self.fc2(x)
4. Early Stopping 策略
当验证集误差连续若干轮无提升时,提前停止训练,避免过拟合。
python
max_patience = 5
patience = 0
best_val_loss = float("inf")
is_early_stop = False
5. RMSNorm 实现与讲解
5.1 原理说明
RMSNorm 是一种替代 LayerNorm 的轻量化归一化方法:
- 不减均值
- 仅用激活值的均方根进行归一化
- 不依赖 batch 维度
数学公式:
RMS ( x ) = 1 n ∑ i = 1 n x i 2 \text{RMS}(x) = \sqrt{\frac{1}{n} \sum_{i=1}^n x_i^2} RMS(x)=n1i=1∑nxi2
RMSNorm ( x ) = x RMS ( x ) + ϵ ⋅ γ \text{RMSNorm}(x) = \frac{x}{\text{RMS}(x) + \epsilon} \cdot \gamma RMSNorm(x)=RMS(x)+ϵx⋅γ
其中 γ \gamma γ 为可学习参数, ϵ \epsilon ϵ 是一个很小的数避免除以 0。
5.2 代码实现
python
class RMSNorm(nn.Module):
def __init__(self, hidden_size, eps=1e-6):
super().__init__()
self.weight = nn.Parameter(torch.ones(hidden_size))
self.eps = eps
def forward(self, x):
rms = torch.sqrt(torch.mean(x ** 2, dim=-1, keepdim=True) + self.eps)
return self.weight * x / rms
5.3 与其他归一化对比
方法 | 是否减均值 | 是否除方差 | 是否依赖 batch |
---|---|---|---|
BatchNorm | 是 | 是 | 是 |
LayerNorm | 是 | 是 | 否 |
RMSNorm | 否 | 是 (仅 RMS) | 否 |
6. 实验建议
- 尝试不同的 Dropout 比例(如 0.1 / 0.3 / 0.5)并观察效果;
- 对比是否每层都加 BatchNorm 是否更优;
- 比较 L2 正则项中 weight decay 的不同取值;
- 使用 RMSNorm 替代 LayerNorm 做对比实验。