以下是基于PyTorch实现TextCNN的代码示例： ### TextCNN 实现 ```python import torch import torch.nn as nn import torch.optim as optim class TextCNN(nn.Module): def __init__(self, vocab_size, embedding_dim, num_filters, filter_sizes, num_classes, dropout_prob): super(TextCNN, self).__init__() # Embedding Layer self.embedding = nn.Embedding(vocab_size, embedding_dim) # Convolutional Layers with different kernel sizes self.convs = nn.ModuleList([ nn.Conv2d(1, num_filters, (fs, embedding_dim)) for fs in filter_sizes ]) # Dropout layer self.dropout = nn.Dropout(dropout_prob) # Fully connected layer self.fc = nn.Linear(len(filter_sizes) * num_filters, num_classes) def forward(self, x): # Input shape: [batch_size, seq_len] embedded_x = self.embedding(x).unsqueeze(1) # Shape: [batch_size, 1, seq_len, embed_dim] conv_outs = [] for conv in self.convs: relu_conv = torch.relu(conv(embedded_x)).squeeze(3) # Shape: [batch_size, num_filters, seq_len - fs + 1] pool = torch.max_pool1d(relu_conv, relu_conv.size(2)).squeeze(2) # Global max pooling over time-step dimension conv_outs.append(pool) concat_features = torch.cat(conv_outs, dim=1) # Concatenate outputs from all filters dropped_features = self.dropout(concat_features) logits = self.fc(dropped_features) # Output logits return logits # Example Usage vocab_size = 10000 # Vocabulary size embedding_dim = 128 # Dimensionality of word embeddings num_filters = 100 # Number of filters per convolutional layer filter_sizes = [3, 4, 5] # Kernel sizes for convolution layers num_classes = 2 # Binary classification task dropout_prob = 0.5 # Dropout probability text_cnn_model = TextCNN( vocab_size=vocab_size, embedding_dim=embedding_dim, num_filters=num_filters, filter_sizes=filter_sizes, num_classes=num_classes, dropout_prob=dropout_prob ) criterion = nn.CrossEntropyLoss() optimizer = optim.Adam(text_cnn_model.parameters(), lr=0.001) print(text_cnn_model) ``` 上述代码定义了一个简单的 `TextCNN` 模型结构，其中包含了嵌入层、卷积层、池化层以及全连接层。模型通过不同的卷积核大小提取局部特征，并最终将这些特征拼接在一起用于分类任务。 --- #### 关键点说明 - **Embedding 层**: 将输入单词索引转换为密集向量表示[^1]。 - **Convolutional Layers**: 使用不同尺寸的卷积核捕获 n-gram 特征[^1]。 - **Pooling**: 对每个通道应用全局最大池化操作以减少维度并保留最重要的特征。 - **Fully Connected Layer**: 输出类别概率分布[^1]。 此实现方式适用于文本分类任务，例如情感分析或主题预测等场景。 --- 问题