Pytorch实现多层LSTM模型，并增加emdedding、Dropout、权重共享等优化

简述

本文是 Pytorch封装简单RNN模型，进行中文训练及文本预测 一文的延申，主要做以下改动：

1.将nn.RNN替换为nn.LSTM，并设置多层LSTM：

既然使用pytorch了，自然不需要手动实现多层，注意nn.RNN和nn.LSTM 在实例化时均有参数num_layers来指定层数，本文设置num_layers=2；

2.新增emdedding层，替换掉原来的nn.functional.one_hot向量化，这样得到的emdedding层可以用来做词向量分布式表示；

3.在emdedding后、LSTM内部、LSTM后均增加Dropout层，来抑制过拟合：

在nn.LSTM内部的Dropout可以通过实例化时的参数dropout来设置，需要注意pytorch仅在两层lstm之间应用Dropout，不会在最后一层的LSTM输出上应用Dropout。

emdedding后、LSTM后与线性层之间则需要手动添加Dropout层。

4.考虑emdedding与最后的Linear层共享权重：

这样做可以在保证精度的情况下，减少学习参数，但本文代码没有实现该部分。

不考虑第四条时，模型结构如下：

代码

模型代码：

class MyLSTM(nn.Module):  
    def __init__(self, vocab_size, wordvec_size, hidden_size, num_layers=2, dropout=0.5):  
        super(MyLSTM, self).__init__()  
        self.vocab_size = vocab_size  
        self.word_vec_size = wordvec_size  
        self.hidden_size = hidden_size  
  
        self.embedding = nn.Embedding(vocab_size, wordvec_size)  
        self.dropout = nn.Dropout(dropout)  
        self.rnn = nn.LSTM(wordvec_size, hidden_size, num_layers=num_layers, dropout=dropout)  
        # self.rnn = rnn_layer  
        self.linear = nn.Linear(self.hidden_size, vocab_size)  
  
    def forward(self, x, h0=None, c0=None):  
        # nn.Embedding 需要的类型 (IntTensor or LongTensor)        # 传过来的X是(batch_size, seq), embedding之后 是(batch_size, seq, vocab_size)  
        # nn.LSTM 支持的X默认为(seq, batch_size, vocab_size)  
        # 若想用(batch_size, seq, vocab_size)作参数, 则需要在创建self.embedding实例时指定batch_first=True  
        # 这里用(seq, batch_size, vocab_size) 作参数，所以先给x转置，再embedding，以便再将结果传给lstm  
        x = x.T  
        x.long()  
        x = self.embedding(x)  
  
        x = self.dropout(x)  
        outputs, (h0, c0) = self.rnn(x, (h0, c0))
  
        outputs = self.dropout(outputs)  
  
        outputs = outputs.reshape(-1, self.hidden_size)  
  
        outputs = self.linear(outputs)  
        return outputs, (h0, c0)  
  
    def init_state(self, device, batch_size=1):  
        return (torch.zeros((self.rnn.num_layers, batch_size, self.hidden_size), device=device),  
                torch.zeros((self.rnn.num_layers, batch_size, self.hidden_size), device=device))

训练代码：

模型应用可以参考 Pytorch封装简单RNN模型，进行中文训练及文本预测 一文。

def start_train():  
    # device = torch.device("cpu")  
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")  
    print(f'\ndevice: {device}')  
  
    corpus, vocab = load_corpus("../data/COIG-CQIA/chengyu_qa.txt")  
  
    vocab_size = len(vocab)  
    wordvec_size = 100  
    hidden_size = 256  
    epochs = 1  
    batch_size = 50  
    learning_rate = 0.01  
    time_size = 4  
    max_grad_max_norm = 0.5  
    num_layers = 2  
    dropout = 0.5  
  
    dataset = make_dataset(corpus=corpus, time_size=time_size)  
    data_loader = data.DataLoader(dataset=dataset, batch_size=batch_size, shuffle=True)  
  
    net = MyLSTM(vocab_size=vocab_size, wordvec_size=wordvec_size, hidden_size=hidden_size, num_layers=num_layers, dropout=dropout)  
    net.to(device)  
  
    # print(net.state_dict())  
  
    criterion = nn.CrossEntropyLoss()  
    criterion.to(device)  
    optimizer = optim.Adam(net.parameters(), lr=learning_rate)  
  
    writer = SummaryWriter('./train_logs')  
    # 随便定义个输入, 好使用add_graph  
    tmp = torch.randint(0, 100, size=(batch_size, time_size)).to(device)  
    h0, c0 = net.init_state(batch_size=batch_size, device=device)  
    writer.add_graph(net, [tmp, h0, c0])  
  
    loss_counter = 0  
    total_loss = 0  
    ppl_list = list()  
    total_train_step = 0  
  
    for epoch in range(epochs):  
        print('------------Epoch {}/{}'.format(epoch + 1, epochs))  
  
        for X, y in data_loader:  
            X, y = X.to(device), y.to(device)  
            # 这里batch_size=X.shape[0]是因为在加载数据时, DataLoader没有设置丢弃不完整的批次, 所以存在实际批次不满足设定的batch_size  
            h0, c0 = net.init_state(batch_size=X.shape[0], device=device)  
            outputs, (hn, cn) = net(X, h0, c0)  
            optimizer.zero_grad()  
            # y也变成 时间序列*批次大小的行数, 才和 outputs 一致  
            y = y.T.reshape(-1)  
            # 交叉熵的第二个参数需要LongTorch  
            loss = criterion(outputs, y.long())  
            loss.backward()  
            # 求完梯度之后可以考虑梯度裁剪, 再更新梯度  
            grad_clipping(net, max_grad_max_norm)  
            optimizer.step()  
  
            total_loss += loss.item()  
            loss_counter += 1  
            total_train_step += 1  
            if total_train_step % 10 == 0:  
                print(f'Epoch: {epoch + 1}, 累计训练次数: {total_train_step}, 本次loss: {loss.item():.4f}')  
                writer.add_scalar('train_loss', loss.item(), total_train_step)  
  
        ppl = np.exp(total_loss / loss_counter)  
        ppl_list.append(ppl)  
        print(f'Epoch {epoch + 1} 结束, batch_loss_average: {total_loss / loss_counter}, perplexity: {ppl}')  
        writer.add_scalar('ppl', ppl, epoch + 1)  
        total_loss = 0  
        loss_counter = 0  
  
        torch.save(net.state_dict(), './save/epoch_{}_ppl_{}.pth'.format(epoch + 1, ppl))  
  
    writer.close()  
    return net, ppl_list