├── README.md ├── datas ├── testdata_for_number.xlsx └── traindata.xlsx ├── images └── use.txt ├── label_lstm_keras.py ├── models └── use.txt └── number_lstm_keras.py /README.md: -------------------------------------------------------------------------------- 1 | # lstm_example(lstm的实例) 2 | >based on keras(基于keras框架) 3 | 4 | ## 文件说明 5 | 6 | - data_file1（数据文件1）:`traindata.xlsx` 7 | 用于回归和分类问题的训练 8 | - data_file2（数据文件2）:`testdata_for_number.xlsx` 9 | 用于回归问题的预测 10 | - code_file1（代码文件1）:`number_lstm_keras.py` 11 | predict value: for **regression** question 12 | 预测数值：**回归问题** 13 | - code_file2（代码文件2）:`label_lstm_keras.py` 14 | predict label: for **classification** question 15 | 预测标签：**分类问题** 16 | 17 | ## 对训练过程可视化 18 | 19 | 对训练过程可视化 20 | 以**分类**问题为例 21 | 22 | ![](https://img-blog.csdnimg.cn/20190320145213663.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM4NDEyODY4,size_16,color_FFFFFF,t_70) 23 | changes of loss during training(训练中loss的变化） 24 | ![](https://img-blog.csdnimg.cn/20190320145317911.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM4NDEyODY4,size_16,color_FFFFFF,t_70) 25 | changes of accuracy during training(训练中acuracy的变化） 26 | ## 更多信息 27 | more detailed contents will be in codes or [my blog](https://blog.csdn.net/qq_38412868/article/details/88688581) 28 | 更多细节详见代码或者我的博客 29 | -------------------------------------------------------------------------------- /datas/testdata_for_number.xlsx: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/fff2zrx/lstm_example/ca8013bcc9e9cb598e1b3c87f0986f5f9d728d7b/datas/testdata_for_number.xlsx -------------------------------------------------------------------------------- /datas/traindata.xlsx: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/fff2zrx/lstm_example/ca8013bcc9e9cb598e1b3c87f0986f5f9d728d7b/datas/traindata.xlsx -------------------------------------------------------------------------------- /images/use.txt: -------------------------------------------------------------------------------- 1 | #储存图片用的文件夹 -------------------------------------------------------------------------------- /label_lstm_keras.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | #利用lstm分类 3 | __author__ = 'fff_zrx' 4 | import pandas as pd 5 | import numpy as np 6 | from numpy import array 7 | import random 8 | from sklearn.preprocessing import OneHotEncoder 9 | from sklearn import preprocessing 10 | from keras.models import Sequential 11 | from keras.models import load_model 12 | from keras.layers import Dense, Dropout 13 | from keras.layers import LSTM 14 | from keras import regularizers 15 | from matplotlib import pyplot 16 | import openpyxl 17 | import os 18 | os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' 19 | # ---- 数据导入 ---- 20 | data = pd.read_excel("./datas/traindata.xlsx") 21 | origin_data_x = data.iloc[:,2:].values 22 | origin_data_y=data.iloc[:,1].values 23 | index = [j for j in range(len(origin_data_x))] 24 | random.shuffle(index) 25 | origin_data_y = origin_data_y[index] 26 | origin_data_x = origin_data_x[index] 27 | # ---- 参数定义---- 28 | split_point=int(len(origin_data_x)*0.8) 29 | input_size=11 30 | time_step =5 31 | labels=5 32 | epochs=200 33 | batch_size=72 34 | # 标准化，工具函数 35 | def normal(data): 36 | mean=np.mean(data,axis=0) 37 | std=np.std(data,axis=0) 38 | return (data-mean)/std 39 | #对labels进行one-hot编码 40 | def label2hot(labels): 41 | values = array(labels) 42 | onehot_encoder = OneHotEncoder(sparse=False) 43 | integer_encoded = values.reshape(len(values), 1) 44 | onehot_encoded = onehot_encoder.fit_transform(integer_encoded) 45 | return onehot_encoded 46 | hot_data_y=label2hot(origin_data_y[:]) 47 | #hot_data_y.append(onehot_encoded) 48 | #hot_data_y=array(hot_data_y).transpose((1,0,2)) 49 | # 训练集数据 50 | train_x= origin_data_x[:split_point] 51 | scaler = preprocessing.StandardScaler().fit(train_x) 52 | train_x=scaler.transform(train_x) 53 | # train_x = normal(train_x) 54 | train_x=train_x.reshape([-1,input_size,time_step]) 55 | train_x=np.transpose(train_x,[0,2,1]) 56 | train_y = hot_data_y[:split_point] 57 | # 测试集数据 58 | test_x= origin_data_x[split_point:] 59 | test_x=scaler.transform(test_x) 60 | # test_x = normal(test_x) 61 | test_x=test_x.reshape([-1,input_size,time_step]) 62 | test_x=np.transpose(test_x,[0,2,1]) 63 | test_y = hot_data_y[split_point:] 64 | print("Data processing is finished!") 65 | # design network 66 | model = Sequential() 67 | # model.add(LSTM(30, input_shape=(train_x.shape[1], train_x.shape[2]),kernel_regularizer=regularizers.l2(0.001),activity_regularizer=regularizers.l1(0.001))) 68 | model.add(LSTM(30, input_shape=(train_x.shape[1], train_x.shape[2]))) 69 | # model.add(Dropout(0.2)) 70 | # model.add(LSTM(6, return_sequences=False)) 71 | model.add(Dense(labels, activation='softmax')) 72 | model.summary() #打印出模型概况 73 | model.compile(loss='categorical_crossentropy', optimizer='adam',metrics=['accuracy']) 74 | # fit network 75 | history = model.fit(train_x,train_y, epochs=epochs, batch_size=batch_size, validation_data=(test_x, test_y), verbose=2, shuffle=True) 76 | #save model after train保存模型文件 77 | model.save('./models/lstm_model_label.h5') 78 | # test the model 79 | score = model.evaluate(test_x, test_y, verbose=2) #evaluate函数按batch计算在某些输入数据上模型的误差 80 | print('Test accuracy:', score[1]) 81 | score = model.evaluate(train_x, train_y, verbose=2) #evaluate函数按batch计算在某些输入数据上模型的误差 82 | print('Train accuracy:', score[1]) 83 | #导出数据 84 | prediction_label = model.predict_classes(test_x) 85 | prediction_label=[i+1 for i in prediction_label] 86 | fact_label=np.argmax(test_y,1) 87 | fact_label=[i+1 for i in fact_label] 88 | analysis=[fact_label, prediction_label] 89 | wb = openpyxl.Workbook() 90 | sheet = wb.active 91 | sheet.title = 'analysis_data' 92 | for i in range(0, 2): 93 | for j in range(0, len(analysis[i])): 94 | sheet.cell(row=j + 1, column=i + 1, value=analysis[i][j]) 95 | wb.save('./datas/analysis_label.xlsx') 96 | print("写入预测数据成功！") 97 | # plot history 98 | pyplot.plot(history.history['loss'], label='train') 99 | pyplot.plot(history.history['val_loss'], label='test') 100 | pyplot.legend() 101 | pyplot.xlabel('Epochs', fontsize = 12) 102 | pyplot.ylabel('Loss', fontsize = 12) 103 | pyplot.savefig("./images/Loss_label.png") 104 | pyplot.show() 105 | pyplot.plot(history.history['acc'], label='train') 106 | pyplot.plot(history.history['val_acc'], label='test') 107 | pyplot.legend() 108 | pyplot.xlabel('Epochs', fontsize = 12) 109 | pyplot.ylabel('Accuracy', fontsize = 12) 110 | pyplot.savefig("./images/Accuracy_label.png") 111 | pyplot.show() 112 | # deletes the existing model 113 | #del model 114 | # load model从模型文件加载模型 115 | #model=load_model('./models/lstm_model_label.h5') 116 | -------------------------------------------------------------------------------- /models/use.txt: -------------------------------------------------------------------------------- 1 | #储存训练完模型文件用的文件夹 -------------------------------------------------------------------------------- /number_lstm_keras.py: -------------------------------------------------------------------------------- 1 | # -*- coding: utf-8 -*- 2 | #利用lstm预测 3 | __author__ = 'fff_zrx' 4 | import pandas as pd 5 | import numpy as np 6 | import random 7 | import os 8 | import keras 9 | import pywt 10 | os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' 11 | from keras.models import Sequential 12 | from keras.models import load_model 13 | from keras.layers import Dense,Dropout 14 | from keras.layers import LSTM 15 | from matplotlib import pyplot 16 | from sklearn import preprocessing 17 | from keras import backend as K 18 | from keras.callbacks import ModelCheckpoint 19 | import openpyxl 20 | from keras import losses 21 | from sklearn.decomposition import PCA 22 | # ---- 数据导入 ---- 23 | data = pd.read_excel("./datas/traindata.xlsx") 24 | origin_data_x = data.iloc[:,2:].values 25 | origin_data_y=data.iloc[:,0].values 26 | index = [j for j in range(len(origin_data_x))] 27 | random.shuffle(index) 28 | origin_data_y = origin_data_y[index] 29 | origin_data_x = origin_data_x[index] 30 | # ---- 参数定义---- 31 | split_point=int(len(origin_data_x)*0.8) 32 | input_size=11 # 输入层维数 33 | time_step =5 # 步长窗口 34 | epochs=150 35 | batch_size=72 36 | # 标准化，工具函数 37 | def calculate_mape(data_x,data_y): 38 | index = list(np.nonzero(data_y)[0]) 39 | data_y = np.array([data_y[i] for i in index]) 40 | predict= model.predict(data_x) 41 | predict = np.array([predict[i] for i in index]) 42 | return np.mean(np.abs(data_y - predict) * std / (np.abs(data_y * std + mean))) 43 | def calculate_mae(data_x, data_y): 44 | index = list(np.nonzero(data_y)[0]) 45 | data_y = np.array([data_y[i] for i in index]) 46 | predict = model.predict(data_x) 47 | predict = np.array([predict[i] for i in index]) 48 | return np.mean(np.abs(data_y - predict) * std) 49 | def mape(y_true, y_pred): 50 | return K.mean(K.abs(y_true - y_pred)*std/(K.abs(y_true*std+mean))) 51 | class mape_callback(keras.callbacks.Callback): 52 | def __init__(self, training_data, validation_data): 53 | self.x = training_data[0] 54 | self.y = training_data[1] 55 | self.x_val = validation_data[0] 56 | self.y_val = validation_data[1] 57 | def on_train_begin(self, logs={}): 58 | return 59 | def on_train_end(self, logs={}): 60 | return 61 | def on_epoch_begin(self, epoch, logs={}): 62 | return 63 | def on_epoch_end(self, epoch, logs={}): 64 | y_pred = self.model.predict(self.x) 65 | mape = np.mean(np.abs(self.y - y_pred)*std/ (np.abs(self.y*std+mean))) 66 | y_pred_val = self.model.predict(self.x_val) 67 | val_mape=np.mean(np.abs(self.y_val - y_pred_val)*std /(np.abs(self.y_val*std+mean))) 68 | print('mape: %s - val_mape: %s' % (mape,val_mape)) 69 | return 70 | def on_batch_begin(self, batch, logs={}): 71 | return 72 | def on_batch_end(self, batch, logs={}): 73 | return 74 | 75 | # 训练集数据 76 | train_x= origin_data_x[:split_point] 77 | scaler = preprocessing.StandardScaler().fit(train_x) 78 | train_x=scaler.transform(train_x) 79 | train_x=train_x.reshape([-1,input_size,time_step]) 80 | train_x=np.transpose(train_x,[0,2,1]) 81 | train_y = origin_data_y[:split_point] 82 | train_y= train_y.reshape([-1,1]) 83 | scaler1 = preprocessing.StandardScaler().fit(train_y) 84 | train_y=scaler1.transform(train_y) 85 | mean=scaler1.mean_ 86 | std=np.sqrt(scaler1.var_) 87 | # 测试集数据 88 | test_x= origin_data_x[split_point:] 89 | test_x=scaler.transform(test_x) 90 | test_x = test_x.reshape([-1, input_size, time_step]) 91 | test_x = np.transpose(test_x, [0, 2, 1]) 92 | test_y = origin_data_y[split_point:] 93 | test_y= test_y.reshape([-1,1]) 94 | test_y=scaler1.transform(test_y) 95 | # design network 96 | model = Sequential() 97 | model.add(LSTM(30, input_shape=(train_x.shape[1], train_x.shape[2]),return_sequences=False)) 98 | # model.add(LSTM(30, return_sequences=False)) 99 | model.add(Dropout(0.5)) 100 | model.add(Dense(1)) 101 | model.summary() #打印出模型概况 102 | model.compile(loss=["mae"], optimizer='adam',metrics=[mape]) 103 | # fit network 104 | # filepath='model_trained.h5' 105 | # # checkpoint = ModelCheckpoint(filepath, monitor='val_loss', verbose=1, save_best_only=True, mode='min') 106 | history = model.fit(train_x, train_y, epochs=epochs, batch_size=batch_size, validation_data=[test_x, test_y],verbose=2, shuffle=True) 107 | #save model after train保存模型文件 108 | model.save('./models/lstm_model_number.h5') 109 | # test the model 110 | print("Testdatasets mape:",calculate_mape(test_x,test_y)) 111 | print("Testdatasets mae:",calculate_mae(test_x,test_y)) 112 | ##plot history画出训练过程 113 | pyplot.figure(1) 114 | pyplot.plot(history.history['loss'], label='train') 115 | pyplot.plot(history.history['val_loss'], label='test') 116 | pyplot.legend() 117 | pyplot.xlabel('Epochs', fontsize = 12) 118 | pyplot.ylabel('Loss', fontsize = 12) 119 | pyplot.savefig("./images/Loss_number.png") 120 | pyplot.show() 121 | pyplot.figure(2) 122 | pyplot.plot(history.history['mape'], label='train') 123 | pyplot.plot(history.history['val_mape'], label='test') 124 | pyplot.legend() 125 | pyplot.xlabel('Epochs', fontsize = 12) 126 | pyplot.ylabel('Mape', fontsize = 12) 127 | pyplot.savefig("./images/Mape_number.png") 128 | pyplot.show() 129 | # deletes the existing model 130 | # del model 131 | # load model 132 | # model=load_model('./models/lstm_model_number.h5') 133 | #---- 待预测数据导入 ---- 134 | data = pd.read_excel("./datas/testdata_for_number.xlsx") 135 | x = data.iloc[:, 4:].values 136 | x = scaler.transform(x) 137 | x = x.reshape([-1, 11, 5]) 138 | x = np.transpose(x, [0, 2, 1]) 139 | y = data.iloc[:, 3].values 140 | true = y.reshape([-1, 1]) 141 | t = data.iloc[:, 2] 142 | # load model 143 | model._make_predict_function() 144 | predict = model.predict(x) 145 | predict = scaler1.inverse_transform(predict) 146 | # 计算mape与mae 147 | index = list(np.nonzero(true)[0]) 148 | after_true = np.array([true[i] for i in index]) 149 | after_predict = np.array([predict[i] for i in index]) 150 | mape = np.mean(np.abs(after_true - after_predict) / (np.abs(after_true))) 151 | mae = np.mean(np.abs(after_true - after_predict)) 152 | print("testday:3月12-15日 Mape:",mape) 153 | print("testday:3月12-15日 Mae:",mae) 154 | y=y.reshape([-1,]) 155 | analysis=[list(y),list(predict)] 156 | mat=np.mat(analysis) 157 | mat=mat.T 158 | np.savetxt('./datas/test_data_number_output.csv',mat,delimiter=',') 159 | print("写入预测数据成功！") 160 | --------------------------------------------------------------------------------