基于keras实现线性回归｜多步预测

 基于keras实现线性回归模型，多步预测问题实质上是多输入单输出的神经网络模型

最近实验中由于自己之前没有学习过tensorflow、pytorch等，无论是自己的实验代码还是对比方法的复现都一头雾水，此篇博客简单记录一下实验的一个小模块线性回归的实现，我的需求是对于输入的训练数据（多维数据）的每一维数据分别用其历史的多步时间数据去预测下个时间戳的数据，并将每一维数据训练所得的模型保存。以下是具体的代码：

import numpy as np
from keras.models import Sequential
from keras.layers import Dense
import pandas as pd
import scipy as sp
from sklearn.metrics import mean_squared_error,mean_absolute_error

# 滑动窗口的实现
def seq2instance(data, P, Q):
    num_step = len(data)
    num_sample = num_step - P - Q + 1
    x = np.zeros(shape = (num_sample, P))
    y = np.zeros(shape = (num_sample, Q))
    for i in range(num_sample):
        x[i] = data[i : i + P]
        y[i] = data[i + P : i + P + Q]
    return x, y

# 线性回归实现
def LinearRegression(model_saver_path='/home/qjy/pythonProj/TP/Model/',
           data_file='/home/qjy/pythonProj/TP/data/Trajectory.csv',
           outputData_path='/home/qjy/pythonProj/TP/data/',
           train_ratio=0.7, val_ratio=0.1, test_ratio=0.2, TIME_STEPS=10, INPUT_SIZE=200, LR_epochs=10, LR_batchSize=128):# default parameter
    # Loading & processing data
    col_dict = {'time_stamps': object}
    for i in range(0, INPUT_SIZE):
        col_dict[str(i)] = int
    dataset = pd.read_csv(data_file + '', header=0, index_col=0, dtype=col_dict)
    values = dataset.values
    # Training set
    train_step = int(values.shape[0] * train_ratio)
    train = np.array(values[:train_step,:])
    # validation set
    valid_step = int(values.shape[0] * val_ratio)
    valid = np.array(values[train_step:train_step + valid_step,:])
    # test set
    test_step = int(values.shape[0] * test_ratio)
    test = np.array(values[-test_step:,:])
    # timeStamps
    idx = dataset.index.values[-test_step + TIME_STEPS:]

    # save prediction
    col_titles = []
    col_titles.append('time_stamps')
    for i in range(0, INPUT_SIZE):
        col_titles.append(str(i))
    df = pd.DataFrame(columns=col_titles)
    df['time_stamps'] = idx
    df.fillna(0, inplace=True)

    # traverse all dims
    for i in range(0, INPUT_SIZE):
        # split into input and outputs
        train_X, train_y = seq2instance(train[:,i], TIME_STEPS, 1)
        train_y = np.squeeze(train_y)
        valid_X, valid_y = seq2instance(valid[:,i], TIME_STEPS, 1)
        valid_y = np.squeeze(valid_y)
        test_X, test_y = seq2instance(test[:,i], TIME_STEPS, 1)
        test_y = np.squeeze(test_y)

        # build model
        model = Sequential()
        model.add(Dense(units=1, input_dim=TIME_STEPS, activation='sigmoid'))

        # select loss function and optimizer
        model.compile(loss='mae', optimizer='sgd', metrics=['accuracy'])

        # train: epcoch, batch_size
        model.fit(train_X, train_y, epochs=LR_epochs, batch_size=LR_batchSize, validation_data=(valid_X, valid_y), verbose=1, shuffle=True)
        # model evaluate
        score = model.evaluate(test_X, test_y, batch_size=LR_batchSize, verbose=1)
        print(score)
        # W, b = model.layers[0].get_weights()
        # print('Weights=', W, '\n biases=', b)

        # save model
        model_file = model_saver_path + 'epoch' + str(LR_epochs) + '_batchSize' + str(LR_batchSize) \
                     + '_windowSize' + str(TIME_STEPS) + '_model.h5'
        model.save(model_file)

        #predict
        Y_pred = model.predict(test_X).astype(int)
        # inverse transform of prediction
        inv_yhat = np.concatenate(Y_pred)
        # inverse transform of real value
        inv_y = test_y

        # metric
        rmse = sp.sqrt(mean_squared_error(inv_y, inv_yhat))
        print('The RMSE:', rmse)
        mae = mean_absolute_error(inv_y, inv_yhat)
        print('The MAE:', mae)


        for t in range(0, len(idx)):
            curTime = idx[t]
            df.loc[df['time_stamps'] == curTime, str(i)] = inv_yhat[t]

    outputData_file = outputData_path + 'LR_predResult.csv'
    df.to_csv(outputData_file, index=False, header=True)


if __name__ == '__main__':
    LinearRegression()