### 使用 PyTorch 搭建 RNN 实现唐诗生成 #### 数据准备与预处理 为了使用循环神经网络（RNN）生成唐诗，首先需要准备好相应的数据集并对其进行必要的预处理。这一步骤通常包括将诗歌中的汉字转换成计算机可以理解的形式。 ```python import torch from collections import Counter def load_data(file_path): with open(file_path, 'r', encoding='utf-8') as f: poems = f.readlines() all_words = ''.join(poems).replace('\n', '') word_counts = Counter(all_words) vocab_size = len(word_counts) idx_to_word = {i: w for i, (w, _) in enumerate(word_counts.items())} word_to_idx = {w: i for i, w in idx_to_word.items()} sequences = [[word_to_idx[char] for char in poem.strip()] for poem in poems] return sequences, word_to_idx, idx_to_word, vocab_size ``` 这段代码展示了如何加载唐诗文件，并创建了一个从字符到索引的映射表`word_to_idx`和反向映射表`idx_to_word`[^3]。 #### 构建模型结构 接下来定义一个简单的单层 LSTM 或 GRU 的类作为基础架构用于后续训练过程： ```python import torch.nn as nn class PoetryGenerator(nn.Module): def __init__(self, vocab_size, embedding_dim=256, hidden_units=512, num_layers=2): super(PoetryGenerator, self).__init__() self.embedding = nn.Embedding(vocab_size, embedding_dim) self.rnn = nn.LSTM(embedding_dim, hidden_units, num_layers=num_layers, batch_first=True) self.fc_out = nn.Linear(hidden_units, vocab_size) def forward(self, x, h0=None, c0=None): embedded = self.embedding(x) output, _ = self.rnn(embedded, (h0, c0)) prediction = self.fc_out(output[:, -1, :]) return prediction ``` 这里采用的是长短期记忆单元(LSTM)，它能够更好地捕捉长时间依赖关系，在自然语言处理任务中有广泛应用价值[^2]。 #### 训练流程设计 有了上述组件之后就可以开始编写完整的训练逻辑了: ```python device = "cuda" if torch.cuda.is_available() else "cpu" model = PoetryGenerator(vocab_size=vocab_size).to(device) criterion = nn.CrossEntropyLoss().to(device) optimizer = torch.optim.Adam(model.parameters(), lr=0.001) for epoch in range(num_epochs): model.train() total_loss = 0. for seqs in dataloader: inputs = seqs[:-1].unsqueeze(0).to(device) targets = seqs[1:].unsqueeze(0).reshape(-1).to(device) optimizer.zero_grad() outputs = model(inputs) loss = criterion(outputs, targets) loss.backward() optimizer.step() total_loss += loss.item() avg_train_loss = total_loss / len(dataloader) print(f'Epoch [{epoch+1}/{num_epochs}], Loss: {avg_train_loss:.4f}') ``` 此部分实现了标准的小批量随机梯度下降算法来进行参数更新；同时利用交叉熵损失函数评估预测结果的质量[^1]。