从 LSTM 自动编码器馈送分类器数据

Question

目标：

为了减少特征，我构建了一个 LSTM 自动编码器。我的计划是对一些输入进行编码，并在将来将其提供给分类器。编码器获取[batch_size, timesteps, features_of_timesteps 形状的数据，但是在编码器部分的输出层中，我只返回[1, timesteps, features_of_timesteps] 形式的最后一个隐藏状态。

class Encoder(nn.Module):
    def __init__(self, input_size, first_layer, second_layer, n_layers):
        super(Encoder, self).__init__()
        self.n_layers = n_layers
        self.encode = nn.Sequential(nn.LSTM(input_size, first_layer, batch_first=True),
                                    getSequence(),
                                    nn.ReLU(True),
                                    nn.LSTM(first_layer, second_layer),
                                    getLast())
        self.decode = nn.Sequential(nn.LSTM(second_layer, first_layer, batch_first=True),
                                    getSequence(),
                                    nn.ReLU(True),
                                    nn.LSTM(first_layer, input_size),
                                    getSequence())

    def forward(self, x):
        x = x.float()
        x = self.encode(x)
        x = x.repeat(batch_size, 1, 1)
        x = self.decode(x)
        return x

担心：

恐怕我的第二个 LSTM 层在模型编码部分的最后一个隐藏状态是在汇总整个批次的同时降低特征维度。这感觉不对，因为我试图将单个时间序列减少为一个较小的向量，而不是将整批时间序列减少为一个向量。我的担心正确吗？

Answer 1

您的代码存在多个问题，为简单起见，我只给您一个定义明确的模型，下面的代码构建了一个 LSTM 自动编码器，它用形状 (batch_size, timesteps, number_of_features_at_each_timesteps) 重建输入：

import torch
from torch import nn

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

class Encoder(nn.Module):
  def __init__(self, seq_len, n_features, embedding_dim=64):
    super(Encoder, self).__init__()

    self.seq_len, self.n_features = seq_len, n_features
    self.embedding_dim, self.hidden_dim = embedding_dim, 2 * embedding_dim

    self.rnn1 = nn.LSTM(
      input_size=n_features,
      hidden_size=self.hidden_dim,
      num_layers=1,
      batch_first=True
    )
    self.rnn2 = nn.LSTM(
      input_size=self.hidden_dim,
      hidden_size=self.embedding_dim,
      num_layers=1,
      batch_first=True
    )

  def forward(self, x):
    x, (_, _) = self.rnn1(x)
    x, (hidden_n, _) = self.rnn2(x)
    return hidden_n

class Decoder(nn.Module):
  def __init__(self, seq_len, input_dim=64, n_features=1):
    super(Decoder, self).__init__()

    self.seq_len, self.input_dim = seq_len, input_dim
    self.hidden_dim, self.n_features = 2 * input_dim, n_features

    self.rnn1 = nn.LSTM(
      input_size=input_dim,
      hidden_size=input_dim,
      num_layers=1,
      batch_first=True
    )
    self.rnn2 = nn.LSTM(
      input_size=input_dim,
      hidden_size=self.hidden_dim,
      num_layers=1,
      batch_first=True
    )
    self.output_layer = nn.Linear(self.hidden_dim, n_features)

  def forward(self, x):
    x = x.repeat(self.seq_len, 1, 1)
    x = x.permute(1, 0, 2)
    x, (hidden_n, cell_n) = self.rnn1(x)
    x, (hidden_n, cell_n) = self.rnn2(x)
    return self.output_layer(x)

class RecurrentAutoencoder(nn.Module):
  def __init__(self, seq_len, n_features, embedding_dim=64):
    super(RecurrentAutoencoder, self).__init__()

    self.encoder = Encoder(seq_len, n_features, embedding_dim).to(device)
    self.decoder = Decoder(seq_len, embedding_dim, n_features).to(device)

  def forward(self, x):
    print("Inputs size:", x.size())
    x = self.encoder(x)
    print("Representation size: ", x.size())
    x = self.decoder(x)
    print("Outputs size: ", x.size())
    return x

batch_n = 5
seq_len = 10
n_features = 3
inputs = torch.randn(batch_n, seq_len, n_features).to(device)

model = RecurrentAutoencoder(seq_len, n_features).to(device)
y = model(inputs)

输出：

Inputs size: torch.Size([5, 10, 3])
Representation size:  torch.Size([1, 5, 64])
Outputs size:  torch.Size([5, 10, 3])

注意表示（即编码器的输出）的形状为(1, batch_size, embedding_dim)