pytorch实现LSTMAutoEncoder与案例

LSTM AutoEncoder简介

基础的AutoEncoder可以参考&＃xff1a;https://blog.csdn.net/weixin_35757704/article/details/118457110

而LSTM AutoEncoder是将原始的全连接变成了LSTM&＃xff0c;然后构造出来的AutoEncoder模型&＃xff0c;输入与输出是一样的数据为最佳

LSTM AutoEncoder 实现

博主发现网上对于LSTM AutoEncoder的版本都不一样&＃xff0c;通常来讲有&＃xff1a;

1. encoder与decoder都是&＃xff1a;lstm
2. encoder是 lstm &＃43; fc ; decoder是 fc &＃43; lstm

以下是两种网络架构&＃xff1a;

基本LSTM AutoEncoder网络结构

这个结构比较简单&＃xff0c;就是encoder的时候过一个lstm&＃xff0c;decoder的时候再过一个lstm

class LstmAutoEncoder(nn.Module):def __init__(self, input_layer&＃61;300, hidden_layer&＃61;100, batch_size&＃61;20):super(LstmAutoEncoder, self).__init__()self.input_layer &＃61; input_layerself.hidden_layer &＃61; hidden_layerself.batch_size &＃61; batch_sizeself.encoder_lstm &＃61; nn.LSTM(self.input_layer, self.hidden_layer, batch_first&＃61;True)self.decoder_lstm &＃61; nn.LSTM(self.hidden_layer, self.input_layer, batch_first&＃61;True)def forward(self, input_x):input_x &＃61; input_x.view(len(input_x), 1, -1)# encoderencoder_lstm, (n, c) &＃61; self.encoder_lstm(input_x,(torch.zeros(1, self.batch_size, self.hidden_layer),torch.zeros(1, self.batch_size, self.hidden_layer)))# decoderdecoder_lstm, (n, c) &＃61; self.decoder_lstm(encoder_lstm,(torch.zeros(1, self.batch_size, self.input_layer),torch.zeros(1, self.batch_size, self.input_layer)))return decoder_lstm.squeeze()


LSTM&＃43;Fc AutoEncoder网络结构

这个网络结构就是&＃xff1a;

• 在encoder的时候过一个lstm&＃xff0c;然后接一个全连接&＃xff0c;最后用relu激活函数&＃xff1b;
• 在decoder的时候先过全连接&＃xff0c;然后用relu的激活函数&＃xff0c;最后接lstm

class LstmFcAutoEncoder(nn.Module):def __init__(self, input_layer&＃61;300, hidden_layer&＃61;100, batch_size&＃61;20):super(LstmFcAutoEncoder, self).__init__()self.input_layer &＃61; input_layerself.hidden_layer &＃61; hidden_layerself.batch_size &＃61; batch_sizeself.encoder_lstm &＃61; nn.LSTM(self.input_layer, self.hidden_layer, batch_first&＃61;True)self.encoder_fc &＃61; nn.Linear(self.hidden_layer, self.hidden_layer)self.decoder_lstm &＃61; nn.LSTM(self.hidden_layer, self.input_layer, batch_first&＃61;True)self.decoder_fc &＃61; nn.Linear(self.hidden_layer, self.hidden_layer)self.relu &＃61; nn.ReLU()def forward(self, input_x):input_x &＃61; input_x.view(len(input_x), 1, -1)# encoderencoder_lstm, (n, c) &＃61; self.encoder_lstm(input_x,# shape: (n_layers, batch, hidden_size)(torch.zeros(1, self.batch_size, self.hidden_layer),torch.zeros(1, self.batch_size, self.hidden_layer)))encoder_fc &＃61; self.encoder_fc(encoder_lstm)encoder_out &＃61; self.relu(encoder_fc)# decoderdecoder_fc &＃61; self.relu(self.decoder_fc(encoder_out))decoder_lstm, (n, c) &＃61; self.decoder_lstm(decoder_fc,(torch.zeros(1, 20, self.input_layer),torch.zeros(1, 20, self.input_layer)))return decoder_lstm.squeeze()


案例代码

import torch
import torch.nn as nn
import torch.utils.data as Datadef get_train_data():"""得到训练数据&＃xff0c;这里使用随机数生成训练数据&＃xff0c;由此导致最终结果并不好"""def get_tensor_from_pd(dataframe_series) -> torch.Tensor:return torch.tensor(data&＃61;dataframe_series.values)import numpy as npimport pandas as pdfrom sklearn import preprocessing# 生成训练数据x并做归一化后&＃xff0c;构造成dataframe格式&＃xff0c;再转换为tensor格式df &＃61; pd.DataFrame(data&＃61;preprocessing.MinMaxScaler().fit_transform(np.random.randint(0, 10, size&＃61;(2000, 300))))y &＃61; pd.Series(np.random.randint(0, 2, 2000))return get_tensor_from_pd(df).float(), get_tensor_from_pd(y).float()class LstmAutoEncoder(nn.Module):def __init__(self, input_layer&＃61;300, hidden_layer&＃61;100, batch_size&＃61;20):super(LstmAutoEncoder, self).__init__()self.input_layer &＃61; input_layerself.hidden_layer &＃61; hidden_layerself.batch_size &＃61; batch_sizeself.encoder_lstm &＃61; nn.LSTM(self.input_layer, self.hidden_layer, batch_first&＃61;True)self.decoder_lstm &＃61; nn.LSTM(self.hidden_layer, self.input_layer, batch_first&＃61;True)def forward(self, input_x):input_x &＃61; input_x.view(len(input_x), 1, -1)# encoderencoder_lstm, (n, c) &＃61; self.encoder_lstm(input_x,(torch.zeros(1, self.batch_size, self.hidden_layer),torch.zeros(1, self.batch_size, self.hidden_layer)))# decoderdecoder_lstm, (n, c) &＃61; self.decoder_lstm(encoder_lstm,(torch.zeros(1, self.batch_size, self.input_layer),torch.zeros(1, self.batch_size, self.input_layer)))return decoder_lstm.squeeze()class LstmFcAutoEncoder(nn.Module):def __init__(self, input_layer&＃61;300, hidden_layer&＃61;100, batch_size&＃61;20):super(LstmFcAutoEncoder, self).__init__()self.input_layer &＃61; input_layerself.hidden_layer &＃61; hidden_layerself.batch_size &＃61; batch_sizeself.encoder_lstm &＃61; nn.LSTM(self.input_layer, self.hidden_layer, batch_first&＃61;True)self.encoder_fc &＃61; nn.Linear(self.hidden_layer, self.hidden_layer)self.decoder_lstm &＃61; nn.LSTM(self.hidden_layer, self.input_layer, batch_first&＃61;True)self.decoder_fc &＃61; nn.Linear(self.hidden_layer, self.hidden_layer)self.relu &＃61; nn.ReLU()def forward(self, input_x):input_x &＃61; input_x.view(len(input_x), 1, -1)# encoderencoder_lstm, (n, c) &＃61; self.encoder_lstm(input_x,# shape: (n_layers, batch, hidden_size)(torch.zeros(1, self.batch_size, self.hidden_layer),torch.zeros(1, self.batch_size, self.hidden_layer)))encoder_fc &＃61; self.encoder_fc(encoder_lstm)encoder_out &＃61; self.relu(encoder_fc)# decoderdecoder_fc &＃61; self.relu(self.decoder_fc(encoder_out))decoder_lstm, (n, c) &＃61; self.decoder_lstm(decoder_fc,(torch.zeros(1, 20, self.input_layer),torch.zeros(1, 20, self.input_layer)))return decoder_lstm.squeeze()if __name__ &＃61;&＃61; &＃39;__main__&＃39;:# 得到数据x, y &＃61; get_train_data()train_loader &＃61; Data.DataLoader(dataset&＃61;Data.TensorDataset(x, y), # 封装进Data.TensorDataset()类的数据&＃xff0c;可以为任意维度batch_size&＃61;20, # 每块的大小shuffle&＃61;True, # 要不要打乱数据 (打乱比较好)num_workers&＃61;2, # 多进程&＃xff08;multiprocess&＃xff09;来读数据)# 建模三件套&＃xff1a;loss&＃xff0c;优化&＃xff0c;epochsmodel &＃61; LstmAutoEncoder() # lstm# model &＃61; LstmFcAutoEncoder() # lstm&＃43;fc模型loss_function &＃61; nn.MSELoss() # lossoptimizer &＃61; torch.optim.Adam(model.parameters(), lr&＃61;0.001) # 优化器epochs &＃61; 150# 开始训练model.train()for i in range(epochs):for seq, labels in train_loader:optimizer.zero_grad()y_pred &＃61; model(seq).squeeze() # 压缩维度&＃xff1a;得到输出&＃xff0c;并将维度为1的去除single_loss &＃61; loss_function(y_pred, seq)# 若想要获得类别&＃xff0c;二分类问题使用四舍五入的方法即可&＃xff1a;print(torch.round(y_pred))single_loss.backward()optimizer.step()print("Train Step:", i, " loss: ", single_loss)# 每20次&＃xff0c;输出一次前20个的结果&＃xff0c;对比一下效果if i % 20 &＃61;&＃61; 0:test_data &＃61; x[:20]y_pred &＃61; model(test_data).squeeze() # 压缩维度&＃xff1a;得到输出&＃xff0c;并将维度为1的去除print("TEST: ", test_data)print("PRED: ", y_pred)print("LOSS: ", loss_function(y_pred, test_data))