pytorch TCN-LSTM

### PyTorch中TCN-LSTM模型的实现

在PyTorch框架下构建TCN-LSTM混合模型，旨在融合时间卷积网络(TCN)与长短期记忆(LSTM)的优势来增强时序数据预测能力。此架构不仅继承了LSTM对于捕捉长时间跨度内变量间关系的能力[^2]，同时也引入了TCN所特有的局部感受野特性以及因果卷积机制。

#### 构建TCN模块
首先定义TCN层作为特征提取器的一部分：

import torch.nn as nn

class Chomp1d(nn.Module):
    def __init__(self, chomp_size):
        super(Chomp1d, self).__init__()
        self.chomp_size = chomp_size
    
    def forward(self, x):
        return x[:, :, :-self.chomp_size].contiguous()

class TemporalBlock(nn.Module):
    def __init__(self, n_inputs, n_outputs, kernel_size, stride, dilation, padding, dropout=0.2):
        super(TemporalBlock, self).__init__()
        conv1 = nn.Conv1d(n_inputs, n_outputs, kernel_size,
                          stride=stride, padding=padding, dilation=dilation)
        chomp1 = Chomp1d(padding)
        relu1 = nn.ReLU()
        dropout1 = nn.Dropout(dropout)

        conv2 = nn.Conv1d(n_outputs, n_outputs, kernel_size,
                          stride=stride, padding=padding, dilation=dilation)
        chomp2 = Chomp1d(padding)
        relu2 = nn.ReLU()
        dropout2 = nn.Dropout(dropout)

        net = nn.Sequential(conv1, chomp1, relu1, dropout1,
                            conv2, chomp2, relu2, dropout2)
        
        downsample = nn.Conv1d(
            n_inputs, n_outputs, 1) if n_inputs != n_outputs else None
        
        self.net = net
        self.downsample = downsample
        self.relu = nn.ReLU()

    def forward(self, x):
        out = self.net(x)
        res = x if self.downsample is None else self.downsample(x)
        return self.relu(out + res)

class TCN(nn.Module):
    def __init__(self, input_channels, output_channels, num_channels, kernel_size=2, dropout=0.2):
        super(TCN, self).__init__()
        layers = []
        num_levels = len(num_channels)
        for i in range(num_levels):
            dilation_size = 2 ** i
            in_channels = input_channels if i == 0 else num_channels[i-1]
            out_channels = num_channels[i]
            layers += [TemporalBlock(in_channels, out_channels, kernel_size, stride=1, dilation=dilation_size,
                                     padding=(kernel_size-1) * dilation_size, dropout=dropout)]

        self.network = nn.Sequential(*layers)

    def forward(self, x):
        return self.network(x)

#### 整合TCN与LSTM形成复合模型
接下来创建一个类用于组合上述TCN组件和标准LSTM单元：

class TCNLSTMModel(nn.Module):
    def __init__(self, tcn_input_dim, lstm_hidden_dim, output_dim, device='cpu'):
        super().__init__()

        # Define the TCN part of the model
        self.tcn = TCN(input_channels=tcn_input_dim, output_channels=lstm_hidden_dim,
                       num_channels=[lstm_hidden_dim]*8, kernel_size=7, dropout=0.2).to(device=device)

        # Define the LSTM layer after extracting features using TCN
        self.lstm = nn.LSTM(lstm_hidden_dim, hidden_size=lstm_hidden_dim//2,
                           batch_first=True).to(device=device)

        # Output fully connected layer to map from LSTM's last state to prediction space
        self.fc_out = nn.Linear(lstm_hidden_dim//2, output_dim).to(device=device)

    def forward(self, x):
        # Pass through temporal convolutional network first
        cnn_output = self.tcn(x.permute(0, 2, 1)).permute(0, 2, 1)

        # Then pass it into an LSTM cell
        lstm_output, _ = self.lstm(cnn_output)

        # Use only the final time step’s output (many-to-one architecture)
        predictions = self.fc_out(lstm_output[:, -1, :])

        return predictions

通过这种方式设计出来的TCN-LSTM模型能够在保持较高计算效率的同时有效应对复杂模式识别挑战，并且特别适用于具有较强自相关性的连续型数值序列分析场景。