### 使用 PyTorch 实现 TextCNN 文本分类 #### 数据准备与预处理 为了实现TextCNN模型，首先需要准备好文本数据并对其进行必要的预处理操作。这通常包括去除停用词、分词以及转换成数值表示形式。 对于文本向量化处理，可以采用Word2Vec方法来获得单词嵌入矩阵[^1]。此过程能够将词语映射到连续空间中的稠密向量上，从而更好地捕捉语义信息。 ```python import torch from gensim.models import Word2Vec # 假设已经有一个经过清洗后的句子列表sentences model = Word2Vec(sentences, size=embedding_dim, window=5, min_count=1, workers=4) def get_embedding_matrix(word_index): vocab_size = len(word_index) + 1 embedding_matrix = np.zeros((vocab_size, embedding_dim)) for word, i in word_index.items(): try: embedding_vector = model[word] if embedding_vector is not None: embedding_matrix[i] = embedding_vector except KeyError: continue return embedding_matrix ``` #### 构建TextCNN网络结构 接下来定义一个简单的TextCNN类，在其中设置卷积层、池化层等组件以完成特征提取工作。这里使用了多个不同大小的过滤器(filter)，以便于从输入序列的不同长度范围内抽取局部模式。 ```python class TextCNN(torch.nn.Module): def __init__(self, vocab_size, embed_size, num_classes, kernel_sizes=[3, 4, 5], num_channels=100): super(TextCNN, self).__init__() # 加载预训练好的word embeddings作为Embedding Layer初始化参数 self.embedding = nn.Embedding(vocab_size, embed_size) self.constant_embedding = nn.Embedding(vocab_size, embed_size) self.convs = nn.ModuleList([ nn.Conv2d(in_channels=1, out_channels=num_channels, kernel_size=(k, embed_size)) for k in kernel_sizes]) self.dropout = nn.Dropout(0.5) self.decoder = nn.Linear(num_channels * len(kernel_sizes), num_classes) def forward(self, inputs): x_embedded = self.embedding(inputs).unsqueeze(1) x_constant = self.constant_embedding(inputs).unsqueeze(1) pooling_outputs = [] for conv in self.convs: relu_output = F.relu(conv(x_embedded)).squeeze(-1) max_pooling_result = F.max_pool1d(relu_output, relu_output.shape[2]).squeeze(-1) pooling_outputs.append(max_pooling_result) concated_features = torch.cat(pooling_outputs, dim=-1) outputs = self.decoder(concated_features) return outputs ``` #### 训练与评估模型 最后一步就是编写循环迭代逻辑来进行批量梯度下降优化，并定期计算验证集上的性能指标如准确率等。同时也可以保存最佳权重用于后续推理阶段。 ```python device = 'cuda' if torch.cuda.is_available() else 'cpu' net = TextCNN(vocab_size=len(tokenizer.word_index)+1, embed_size=embed_size, num_classes=n_class).to(device) loss_fn = nn.CrossEntropyLoss() optimizer = optim.Adam(net.parameters(), lr=learning_rate) for epoch in range(num_epochs): net.train() running_loss = 0. correct_predictions = 0 for batch_idx, (inputs, labels) in enumerate(train_loader): optimizer.zero_grad() outputs = net(inputs.to(device)) loss = loss_fn(outputs.cpu(), labels.long()) loss.backward() optimizer.step() _, predicted = torch.max(outputs.data, 1) correct_predictions += (predicted == labels).sum().item() running_loss += loss.item() train_accuracy = float(correct_predictions)/len(train_dataset)*100. print(f'Training Accuracy at Epoch {epoch}: %.2f%%'%train_accuracy) ```