CNN、RNN、Transformer 三大架构实战：图像、时序、文本任务代码对比

CNN、RNN、Transformer 三大架构实战图像、时序、文本任务代码对比深度学习领域的三驾马车——卷积神经网络CNN、循环神经网络RNN和Transformer架构各自在特定数据类型上展现出独特优势。本文将带您深入实战通过PyTorch代码对比这三种架构在图像分类、股票预测和情感分析任务中的实现差异揭示模型设计背后的核心思想与技术细节。1. 环境准备与数据加载在开始构建模型前我们需要配置统一的开发环境。推荐使用Python 3.8和PyTorch 1.12环境所有示例代码均可在Colab或本地GPU服务器上运行。import torch import torch.nn as nn import torch.optim as optim from torch.utils.data import Dataset, DataLoader from torchvision import transforms import numpy as np import pandas as pd from sklearn.model_selection import train_test_split # 检查GPU可用性 device torch.device(cuda if torch.cuda.is_available() else cpu) print(fUsing device: {device})1.1 数据集准备我们将使用三种典型数据集图像数据CIFAR-1010类物体分类时序数据Yahoo股票价格数据集文本数据IMDb电影评论情感分析# 图像数据加载示例 transform transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)) ]) # 时序数据预处理函数 def create_sequences(data, seq_length): sequences [] targets [] for i in range(len(data)-seq_length): seq data[i:iseq_length] label data[iseq_length] sequences.append(seq) targets.append(label) return np.array(sequences), np.array(targets) # 文本数据分词处理 from torchtext.data import get_tokenizer tokenizer get_tokenizer(basic_english)2. CNN图像分类实战卷积神经网络通过局部感受野和权值共享特性成为处理网格状数据如图像的首选架构。下面我们实现一个经典的CNN分类器。2.1 模型架构设计class CNNClassifier(nn.Module): def __init__(self, num_classes10): super(CNNClassifier, self).__init__() self.features nn.Sequential( nn.Conv2d(3, 64, kernel_size3, padding1), nn.ReLU(inplaceTrue), nn.MaxPool2d(kernel_size2, stride2), nn.Conv2d(64, 128, kernel_size3, padding1), nn.ReLU(inplaceTrue), nn.MaxPool2d(kernel_size2, stride2), nn.Conv2d(128, 256, kernel_size3, padding1), nn.ReLU(inplaceTrue), nn.Conv2d(256, 256, kernel_size3, padding1), nn.ReLU(inplaceTrue), nn.MaxPool2d(kernel_size2, stride2), ) self.classifier nn.Sequential( nn.Linear(256 * 4 * 4, 1024), nn.ReLU(inplaceTrue), nn.Dropout(p0.5), nn.Linear(1024, num_classes), ) def forward(self, x): x self.features(x) x torch.flatten(x, 1) x self.classifier(x) return x关键组件解析Conv2d二维卷积层提取局部特征MaxPool2d下采样操作增强平移不变性ReLU非线性激活函数Dropout正则化防止过拟合2.2 训练流程对比CNN的训练过程强调数据增强和批量归一化def train_cnn(model, train_loader, criterion, optimizer, epochs10): model.train() for epoch in range(epochs): running_loss 0.0 for i, (inputs, labels) in enumerate(train_loader): inputs, labels inputs.to(device), labels.to(device) # 数据增强 augmented_inputs augment_images(inputs) optimizer.zero_grad() outputs model(augmented_inputs) loss criterion(outputs, labels) loss.backward() optimizer.step() running_loss loss.item() print(fEpoch {epoch1}, Loss: {running_loss/len(train_loader)})提示图像数据通常需要应用随机裁剪、水平翻转等增强技术这能显著提升模型泛化能力。3. RNN时序预测实战循环神经网络通过隐状态记忆历史信息特别适合处理序列数据。我们以股票价格预测为例展示LSTM的实现。3.1 LSTM模型构建class StockPredictor(nn.Module): def __init__(self, input_size1, hidden_size64, num_layers2): super(StockPredictor, self).__init__() self.hidden_size hidden_size self.num_layers num_layers self.lstm nn.LSTM(input_size, hidden_size, num_layers, batch_firstTrue) self.fc nn.Linear(hidden_size, 1) def forward(self, x): h0 torch.zeros(self.num_layers, x.size(0), self.hidden_size).to(device) c0 torch.zeros(self.num_layers, x.size(0), self.hidden_size).to(device) out, _ self.lstm(x, (h0, c0)) out self.fc(out[:, -1, :]) return out时序数据处理要点序列标准化使用滑动窗口归一化教师强制训练时使用真实值作为下一步输入多步预测递归预测或序列到序列架构3.2 差异化的训练策略def train_rnn(model, train_loader, criterion, optimizer, epochs50): model.train() for epoch in range(epochs): for seq, targets in train_loader: seq, targets seq.float().to(device), targets.float().to(device) optimizer.zero_grad() outputs model(seq.unsqueeze(-1)) loss criterion(outputs, targets.unsqueeze(-1)) # 梯度裁剪防止爆炸 nn.utils.clip_grad_norm_(model.parameters(), max_norm1.0) loss.backward() optimizer.step() if epoch % 10 0: print(fEpoch {epoch}, Loss: {loss.item():.4f})RNN训练特有的技术梯度裁剪缓解梯度爆炸问题序列采样处理变长序列双向结构捕获前后文信息4. Transformer文本分类实战Transformer凭借自注意力机制在自然语言处理领域取得突破。我们实现一个简化版的Transformer用于情感分析。4.1 模型架构创新class TextTransformer(nn.Module): def __init__(self, vocab_size, embed_dim, num_heads, num_layers, num_classes): super(TextTransformer, self).__init__() self.embedding nn.Embedding(vocab_size, embed_dim) self.pos_encoder PositionalEncoding(embed_dim) encoder_layer nn.TransformerEncoderLayer( d_modelembed_dim, nheadnum_heads ) self.transformer_encoder nn.TransformerEncoder( encoder_layer, num_layersnum_layers ) self.classifier nn.Linear(embed_dim, num_classes) def forward(self, x): x self.embedding(x) x self.pos_encoder(x) x self.transformer_encoder(x) x x.mean(dim1) # 全局平均池化 return self.classifier(x) class PositionalEncoding(nn.Module): def __init__(self, d_model, max_len5000): super(PositionalEncoding, self).__init__() pe torch.zeros(max_len, d_model) position torch.arange(0, max_len, dtypetorch.float).unsqueeze(1) div_term torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model)) pe[:, 0::2] torch.sin(position * div_term) pe[:, 1::2] torch.cos(position * div_term) self.register_buffer(pe, pe) def forward(self, x): x x self.pe[:x.size(1), :] return xTransformer核心优势自注意力机制动态计算特征重要性位置编码替代RNN的顺序处理能力并行计算显著提升训练效率4.2 文本处理流程def preprocess_text(text, vocab, max_length256): tokens tokenizer(text) indexed [vocab[token] if token in vocab else vocab[unk] for token in tokens] if len(indexed) max_length: indexed [vocab[pad]] * (max_length - len(indexed)) else: indexed indexed[:max_length] return torch.tensor(indexed) # 示例词表构建 from collections import Counter def build_vocab(texts, max_size20000): counter Counter() for text in texts: counter.update(tokenizer(text)) vocab {pad: 0, unk: 1} vocab.update({token: i2 for i, (token, _) in enumerate(counter.most_common(max_size))}) return vocab5. 三大架构对比分析通过实际代码实现我们可以总结出三种架构的关键差异特性CNNRNNTransformer核心机制局部卷积池化循环连接自注意力数据依赖局部相关性时序依赖全局依赖并行能力完全并行时间步串行完全并行典型应用图像分类、目标检测时间序列预测、语音识别机器翻译、文本分类内存效率中等较低较低长序列时参数共享空间共享时间共享无显式共享性能优化建议CNN使用深度可分离卷积减少参数量RNN采用GRU单元平衡效果与效率Transformer使用线性注意力降低复杂度6. 跨架构迁移技巧虽然三种架构设计初衷不同但现代深度学习实践中常出现架构融合class HybridModel(nn.Module): CNNTransformer混合架构示例 def __init__(self): super().__init__() self.cnn nn.Sequential( nn.Conv2d(3, 32, 3), nn.ReLU(), nn.MaxPool2d(2), nn.Conv2d(32, 64, 3), nn.ReLU(), nn.MaxPool2d(2) ) self.transformer nn.TransformerEncoderLayer( d_model64, nhead4 ) self.classifier nn.Linear(64, 10) def forward(self, x): x self.cnn(x) # [B, C, H, W] x x.flatten(2).permute(2, 0, 1) # [L, B, C] x self.transformer(x) x x.mean(dim0) return self.classifier(x)实际项目中根据数据特性选择架构图像文本CNN处理图像Transformer处理文本视频分析3D CNN提取空间特征LSTM建模时序多模态各模态专用编码器跨模态注意力