### CNN 结合 Transformer 的分类任务代码 以下是结合卷积神经网络 (Convolutional Neural Network, CNN) 和变压器模型 (Transformer) 进行图像分类的任务代码。此代码包含详细的注释以便于理解。 ```python import torch import torch.nn as nn import torchvision.models as models class ConvTransformerModel(nn.Module): """ 定义一个结合CNN和Transformer的分类模型。 """ def __init__(self, num_classes=10, cnn_model='resnet18', transformer_dim=512, num_heads=8, num_layers=2): super(ConvTransformerModel, self).__init__() # 使用预训练的ResNet作为特征提取器 if cnn_model == 'resnet18': resnet = models.resnet18(pretrained=True) modules = list(resnet.children())[:-2] # 去掉最后两层（全局平均池化和全连接） self.cnn_features = nn.Sequential(*modules) cnn_out_size = 512 elif cnn_model == 'resnet50': # 如果需要更复杂的CNN结构可以替换为ResNet50 resnet = models.resnet50(pretrained=True) modules = list(resnet.children())[:-2] self.cnn_features = nn.Sequential(*modules) cnn_out_size = 2048 # 将CNN输出转换为适合Transformer输入的形式 self.flatten = nn.Flatten(start_dim=2) self.linear_projection = nn.Linear(cnn_out_size, transformer_dim) # 构建Transformer编码器 encoder_layer = nn.TransformerEncoderLayer(d_model=transformer_dim, nhead=num_heads) self.transformer_encoder = nn.TransformerEncoder(encoder_layer, num_layers=num_layers) # 输出层用于分类 self.fc = nn.Linear(transformer_dim, num_classes) def forward(self, x): """ 前向传播函数。 :param x: 输入张量，形状为(batch_size, channels, height, width) :return: 分类预测结果 """ batch_size = x.size(0) # 提取CNN特征图 cnn_output = self.cnn_features(x) # 形状变为(batch_size, cnn_out_size, H/32, W/32)，H/W取决于输入大小 cnn_output = cnn_output.permute(0, 2, 3, 1).contiguous() # 调整维度顺序至(batch_size, H/32, W/32, cnn_out_size)