train_encoder_decoder.py

train_encoder_decoder.py

from __future__ import print_function #为了确保代码同时兼容Python 2和Python 3版本中的print函数

# 导入标准库和第三方库
import os.path #导入了Python的os.path模块，用于处理文件和目录路径
from os import path #从os模块中导入了path子模块，可以直接使用path来调用os.path中的函数

import sys #导入了sys模块，用于系统相关的参数和函数
import math #导入了math模块，提供了数学运算函数
import numpy as np #导入了NumPy库，并使用np作为别名，NumPy是用于科学计算的基础库
import pandas as pd #导入了Pandas库，并使用pd作为别名，Pandas是用于数据分析的强大库

# 导入深度学习相关库
import tensorflow as tf #导入了TensorFlow深度学习框架

from keras import backend as K #导入了Keras的backend模块，并使用K作为别名，用于访问后端引擎的函数
from keras.models import Model #从Keras导入了Model类，用于定义神经网络模型
from keras.layers import LSTM, GRU, TimeDistributed, Input, Dense, RepeatVector #从Keras导入了LSTM、Input和Dense等神经网络层
from keras.callbacks import CSVLogger, EarlyStopping, TerminateOnNaN #从Keras导入了CSVLogger、EarlyStopping和TerminateOnNaN等回调函数，用于模型训练时的控制和记录
from keras import regularizers #从Keras导入了regularizers模块，用于正则化
from keras.optimizers import Adam #从Keras导入了Adam优化器，用于编译模型时指定优化算法

# 导入其他功能模块
from functools import partial, update_wrapper #从Python标准库functools中导入了partial和update_wrapper函数，用于函数式编程中的功能扩展和包装

# 这个函数的作用是创建一个部分应用（partial application）的函数，并保留原始函数的文档字符串等信息。
def wrapped_partial(func, *args, **kwargs):
    partial_func = partial(func, *args, **kwargs)
    update_wrapper(partial_func, func)
    return partial_func

# 这是一个自定义的损失函数，计算加权的均方误差（Mean Squared Error），其中y_true是真实值，y_pred是预测值，weights是权重。
def weighted_mse(y_true, y_pred, weights):
    return K.mean(K.square(y_true - y_pred) * weights, axis=-1)

# 这部分代码用于选择使用的GPU设备。它从命令行参数中获取一个整数值gpu，如果gpu小于3，则设置CUDA环境变量以指定使用的GPU设备
import os
gpu = int(sys.argv[-13])
if gpu < 3:
    os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"   # see issue #152
    os.environ["CUDA_VISIBLE_DEVICES"]= "{}".format(gpu)

    from tensorflow.python.client import device_lib
    print(device_lib.list_local_devices())

# 这部分代码获取了一系列命令行参数，并将它们分别赋值给变量 
# 这些参数可能包括数据集名称、训练的批次数量、训练周期数、学习率、正则化惩罚、Dropout率、耐心（用于Early Stopping）等 
imp = sys.argv[-1]
T = sys.argv[-2]
t0 = sys.argv[-3]
dataname = sys.argv[-4] 
nb_batches = sys.argv[-5]
nb_epochs = sys.argv[-6]
lr = float(sys.argv[-7])
penalty = float(sys.argv[-8])
dr = float(sys.argv[-9])
patience = sys.argv[-10]
n_hidden = int(sys.argv[-11])
hidden_activation = sys.argv[-12]

# results_directory 是一个字符串，表示将要创建的结果文件夹路径。dataname 是之前从命令行参数中获取的数据集名称
# 如果这个文件夹路径不存在，就使用 os.makedirs 函数创建它。这个路径通常用于存储训练模型的结果或者日志
results_directory = 'results/encoder-decoder/{}'.format(dataname)

if not os.path.exists(results_directory):
    os.makedirs(results_directory)

# 定义了一个函数 create_model，用于创建、编译和返回一个循环神经网络（RNN）模型
def create_model(n_pre, n_post, nb_features, output_dim, lr, penalty, dr, n_hidden, hidden_activation):
    """ 
        creates, compiles and returns a RNN model 
        @param nb_features: the number of features in the model
    """
    
    # 这里定义了两个输入层：inputs 是一个形状为 (n_pre, nb_features) 的输入张量，用于模型的主输入；weights_tensor 是一个形状相同的张量，用于传递权重或其他需要的信息
    inputs = Input(shape=(n_pre, nb_features), name="Inputs")  
    weights_tensor = Input(shape=(n_pre, nb_features), name="Weights") 
    
   # 这里使用了两个 LSTM 层：lstm_1 是一个具有 n_hidden 个单元的 LSTM 层，应用了 dropout 和 recurrent_dropout，并且返回整个时间序列的输出。lstm_2 是一个相同的 LSTM 层，但它只返回最后一个时间步的输出。
    lstm_1 = LSTM(n_hidden, dropout=dr, recurrent_dropout=dr, activation=hidden_activation, return_sequences=True, name='LSTM_1')(inputs) # Encoder
    lstm_2 = LSTM(n_hidden, activation=hidden_activation, return_sequences=False, name='LSTM_2')(lstm_1) # Encoder
    repeat = RepeatVector(n_post, name='Repeat')(lstm_2) # get the last output of the LSTM and repeats it
    gru_1 = GRU(n_hidden, activation=hidden_activation, return_sequences=True, name='Decoder')(repeat)  # Decoder
    output= TimeDistributed(Dense(output_dim, activation='linear', kernel_regularizer=regularizers.l2(penalty), name='Dense'), name='Outputs')(gru_1)

    model = Model([inputs, weights_tensor], output)

    # Compile
    cl = wrapped_partial(weighted_mse, weights=weights_tensor)
    model.compile(optimizer=Adam(lr=lr), loss=cl)

    print(model.summary()) 

    return model

def train_model(model, dataX, dataY, weights, nb_epoches, nb_batches):

    # Prepare model checkpoints and callbacks

    stopping = EarlyStopping(monitor='val_loss', patience=int(patience), min_delta=0, verbose=1, mode='min', restore_best_weights=True)

    csv_logger = CSVLogger('results/encoder-decoder/{}/training_log_{}_{}_{}_{}_{}_{}_{}_{}.csv'.format(dataname,dataname,imp,hidden_activation,n_hidden,patience,dr,penalty,nb_batches), separator=',', append=False)

    terminate = TerminateOnNaN()

    # Model fit
    
    history = model.fit(x=[dataX,weights], 
        y=dataY, 
        batch_size=nb_batches, 
        verbose=1,
        epochs=nb_epoches, 
        callbacks=[stopping,csv_logger,terminate],
        validation_split=0.2)

def test_model():

    n_post = int(1)
    n_pre =int(t0)-1
    seq_len = int(T)

    wx = np.array(pd.read_csv("data/{}-wx-{}.csv".format(dataname,imp)))

    print('raw wx shape', wx.shape)  

    wXC = []
    for i in range(seq_len-n_pre-n_post):
        wXC.append(wx[i:i+n_pre]) 
   
    wXC = np.array(wXC)

    print('wXC shape:', wXC.shape)

    x = np.array(pd.read_csv("data/{}-x-{}.csv".format(dataname,imp)))

    print('raw x shape', x.shape) 

    dXC, dYC = [], []
    for i in range(seq_len-n_pre-n_post):
        dXC.append(x[i:i+n_pre])
        dYC.append(x[i+n_pre:i+n_pre+n_post])

    dataXC = np.array(dXC)
    dataYC = np.array(dYC)
    
    print('dataXC shape:', dataXC.shape)
    print('dataYC shape:', dataYC.shape)

    nb_features = dataXC.shape[2]
    output_dim = dataYC.shape[2]

    # create and fit the encoder-decoder network
    print('creating model...')
    model = create_model(n_pre, n_post, nb_features, output_dim, lr, penalty, dr, n_hidden, hidden_activation)

    train_model(model, dataXC, dataYC, wXC, int(nb_epochs), int(nb_batches))

    # now test

    print('Generate predictions on full training set')

    preds_train = model.predict([dataXC,wXC], batch_size=int(nb_batches), verbose=1)

    print('predictions shape =', preds_train.shape)

    preds_train = np.squeeze(preds_train)

    print('predictions shape (squeezed)=', preds_train.shape)

    print('Saving to results/encoder-decoder/{}/encoder-decoder-{}-train-{}-{}-{}-{}-{}-{}.csv'.format(dataname,dataname,imp,hidden_activation,n_hidden,patience,dr,penalty,nb_batches))

    np.savetxt("results/encoder-decoder/{}/encoder-decoder-{}-train-{}-{}-{}-{}-{}-{}.csv".format(dataname,dataname,imp,hidden_activation,n_hidden,patience,dr,penalty,nb_batches), preds_train, delimiter=",")

    print('Generate predictions on test set')
    
    wy = np.array(pd.read_csv("data/{}-wy-{}.csv".format(dataname,imp)))

    print('raw wy shape', wy.shape)  

    wY = []
    for i in range(seq_len-n_pre-n_post):
        wY.append(wy[i:i+n_pre]) # weights for outputs
    
    wXT = np.array(wY)

    print('wXT shape:', wXT.shape)

    y = np.array(pd.read_csv("data/{}-y-{}.csv".format(dataname,imp)))

    print('raw y shape', y.shape)  

    dXT = []
    for i in range(seq_len-n_pre-n_post):
        dXT.append(y[i:i+n_pre]) # treated is input

    dataXT = np.array(dXT)

    print('dataXT shape:', dataXT.shape)
  
    preds_test = model.predict([dataXT, wXT], batch_size=int(nb_batches), verbose=1)

    print('predictions shape =', preds_test.shape)
    
    preds_test = np.squeeze(preds_test)

    print('predictions shape (squeezed)=', preds_test.shape)

    print('Saving to results/encoder-decoder/{}/encoder-decoder-{}-test-{}-{}-{}-{}-{}-{}.csv'.format(dataname,dataname,imp,hidden_activation,n_hidden,patience,dr,penalty,nb_batches))

    np.savetxt("results/encoder-decoder/{}/encoder-decoder-{}-test-{}-{}-{}-{}-{}-{}.csv".format(dataname,dataname,imp,hidden_activation,n_hidden,patience,dr,penalty,nb_batches), preds_test, delimiter=",")

def main():
    test_model()
    return 1

if __name__ == "__main__":
    main()