当前位置：首页 > java >正文

[训练和优化] 3. 模型优化

java 2025/8/26 20:16:50

👋 你好！这里有实用干货与深度分享✨✨ 若有帮助，欢迎：
👍 点赞 | ⭐ 收藏 | 💬 评论 | ➕ 关注，解锁更多精彩！
📁 收藏专栏即可第一时间获取最新推送🔔。
📖后续我将持续带来更多优质内容，期待与你一同探索知识，携手前行，共同进步🚀。

人工智能

模型优化

本文详细介绍深度学习模型的优化技术，包括正则化、梯度裁剪、早停、模型集成等方法，帮助提升模型性能和泛化能力。

1. 正则化方法

1.1 权重正则化

通过L1/L2正则化抑制模型复杂度，防止过拟合。

import torchclass L1L2Regularizer:def __init__(self, l1_lambda=0.0, l2_lambda=0.0):self.l1_lambda = l1_lambdaself.l2_lambda = l2_lambdadef __call__(self, model):reg_loss = 0for param in model.parameters():if param.requires_grad:# L1正则化reg_loss += self.l1_lambda * torch.sum(torch.abs(param))# L2正则化reg_loss += self.l2_lambda * torch.sum(param ** 2)return reg_loss# 使用示例
regularizer = L1L2Regularizer(l1_lambda=1e-5, l2_lambda=1e-4)
reg_loss = regularizer(model)
total_loss = task_loss + reg_loss

1.2 Dropout实现

Dropout可有效缓解过拟合，提升模型泛化能力。

import torch
import torch.nn as nnclass CustomDropout(nn.Module):def __init__(self, p=0.5, training=True):super().__init__()self.p = pself.training = trainingdef forward(self, x):if not self.training or self.p == 0:return xmask = torch.bernoulli(torch.ones_like(x) * (1 - self.p))return x * mask / (1 - self.p)# 在模型中使用
self.dropout = CustomDropout(p=0.5)

2. 梯度处理

2.1 梯度裁剪

防止梯度爆炸，提升训练稳定性。

import torchdef clip_gradients(model, clip_value=1.0, clip_norm=None):if clip_norm is not None:torch.nn.utils.clip_grad_norm_(model.parameters(), clip_norm)else:torch.nn.utils.clip_grad_value_(model.parameters(), clip_value)def train_with_gradient_clipping(model, train_loader, criterion,optimizer, device, clip_value=1.0):model.train()for data, target in train_loader:data, target = data.to(device), target.to(device)optimizer.zero_grad()output = model(data)loss = criterion(output, target)loss.backward()# 应用梯度裁剪clip_gradients(model, clip_value)optimizer.step()

2.2 梯度累积

节省显存，模拟大批量训练。

class GradientAccumulator:def __init__(self, model, accumulation_steps):self.model = modelself.accumulation_steps = accumulation_stepsself.current_step = 0def step(self, loss):# 缩放损失loss = loss / self.accumulation_stepsloss.backward()self.current_step += 1return self.current_step % self.accumulation_steps == 0def reset(self):self.current_step = 0

3. 早停策略

3.1 验证集早停

防止过拟合，自动停止训练并保存最佳模型。

class EarlyStopping:def __init__(self, patience=7, min_delta=0):self.patience = patienceself.min_delta = min_deltaself.counter = 0self.best_loss = Noneself.early_stop = Falsedef __call__(self, val_loss):if self.best_loss is None:self.best_loss = val_losselif val_loss > self.best_loss - self.min_delta:self.counter += 1if self.counter >= self.patience:self.early_stop = Trueelse:self.best_loss = val_lossself.counter = 0return self.early_stop# 使用示例
early_stopping = EarlyStopping(patience=10)
for epoch in range(num_epochs):train_loss = train_epoch(model, train_loader, criterion, optimizer)val_loss = validate(model, val_loader, criterion)if early_stopping(val_loss):print('Early stopping triggered')break

4. 模型集成

4.1 模型平均

集成多个模型预测结果，提升鲁棒性和准确率。

import torchclass ModelEnsemble:def __init__(self, models):self.models = modelsdef predict(self, x):predictions = []for model in self.models:model.eval()with torch.no_grad():pred = model(x)predictions.append(pred)# 对预测结果取平均return torch.mean(torch.stack(predictions), dim=0)# 使用示例
models = [train_model() for _ in range(5)]  # 训练多个模型
ensemble = ModelEnsemble(models)
prediction = ensemble.predict(test_data)

4.2 权重平均

直接对模型参数加权平均，获得更稳健的模型。

import copydef average_model_weights(models):"""平均多个模型的权重"""avg_model = copy.deepcopy(models[0])avg_dict = avg_model.state_dict()for key in avg_dict.keys():# 初始化为第一个模型的权重avg_dict[key] = avg_dict[key].clone()# 累加其他模型的权重for model in models[1:]:avg_dict[key] += model.state_dict()[key]# 计算平均值avg_dict[key] = avg_dict[key] / len(models)avg_model.load_state_dict(avg_dict)return avg_model