############導入必要的庫######### import pandas as pd import numpy as np import matplotlib.pyplot as plt from sklearn.preprocessing import MinMaxScaler from keras.models import Sequential from keras.layers import Dense, LSTM, Dropout, GRU from keras.optimizers import SGD在這里插入代碼片

數據的加載

## 將date這一列設為索引 dataset = pd.read_csv("datasets_8388_11883_IBM_2006-01-01_to_2018-01-01.csv", index_col='Date', parse_dates=['Date']) dataset.head() #顯示的是前10列 dataset.shape #獲取的是數據集的維度 >>(3020, 6) # 檢查數據是否有缺失值 # 獲取了其中的一列當做訓練集與測試集 train_set = dataset[:'2016'].iloc[:, 1:2].values # 訓練集，獲取到第一列的數據 test_set = dataset['2017':].iloc[:,1:2].values # 測試集，獲取到第一列的數據 train_set.shape >>(2769, 1) test_set.shape >>(251, 1)

定義顯示函數

def plot_predictions(test_result, predict_restult):"""test_result: 真實值predict_result: 預測值"""plt.plot(test_result, color='red', label='IBM True Stock Price')plt.plot(predict_restult, color='blue', label="IMB prdicted Stock Price")plt.title("IBM Stock Price")plt.xlabel("Time")plt.ylabel("Stock Price")plt.legend() # 給圖加上圖例plt.show()

繪制測試集與訓練集的數目

dataset['High'][:"2016"].plot(figsize=(16, 4), legend=True) dataset['High']["2017":].plot(figsize=(16, 4), legend=True) plt.title("IBM Stock Price") plt.legend(['Train set(before 2016)', 'Test set(after 2017)']) plt.show()

如何制作數據將數據喂到模型里

# 正則化：將每一維的特征映射到指定的區間：【0，1】 sc = MinMaxScaler(feature_range=[0, 1]) train_set_scaled = sc.fit_transform(train_set) # 創建序列數據集（訓練和測試） # 這個代碼的意思就是制作一個時間序列的預處理 # 60個時間步為一個樣本，1個輸出 X_train = [] ##這里是列表 y_train = [] for i in range(60, 2769):X_train.append(train_set_scaled[i-60:i, 0]) y_train.append(train_set_scaled[i, 0])X_train, y_train = np.array(X_train), np.array(y_train) # numpy類型變換 X_train.shape >>>(2709, 60) X_train[0] >>>array([0.06065089, 0.06029868, 0.06311637, 0.0781206 , 0.07262609,0.07171034, 0.07657087, 0.07058326, 0.0669907 , 0.06494787,0.075796 , 0.07361229, 0.06417301, 0.05621302, 0.05783319,0.05409975, 0.05431107, 0.05515638, 0.05543815, 0.05677656,0.05846717, 0.05388842, 0.04811214, 0.04233587, 0.04402649,0.0490279 , 0.04832347, 0.05297267, 0.05614258, 0.05290223,0.05325444, 0.04909834, 0.04994365, 0.04797126, 0.05431107,0.05212736, 0.04726684, 0.04895745, 0.04656241, 0.04839391,0.04416737, 0.0485348 , 0.04719639, 0.04825303, 0.05395886,0.05663567, 0.05853762, 0.05959425, 0.06375035, 0.06917442,0.06889265, 0.06670893, 0.06910397, 0.07783883, 0.07565511,0.07276698, 0.06889265, 0.0656523 , 0.06656805, 0.06769513]) ### lstm的輸入模塊 # LSTM的輸入：(samples, sequence_length, features) # reshape: 訓練集(2709,60) ---> (2709, 60, 1) X_train = np.reshape(X_train, (X_train.shape[0], X_train.shape[1], 1)) X_train.shape >>> (2709, 60, 1)

構建lstm網絡

model = Sequential() # LSTM 第一層 ##輸入的是步數與維度 model.add(LSTM(128, return_sequences=True, input_shape=(X_train.shape[1], 1))) model.add(Dropout(0.2))# LSTM 第二層 model.add(LSTM(128, return_sequences=True)) model.add(Dropout(0.2))# LSTM 第三層 model.add(LSTM(128)) model.add(Dropout(0.2))# Dense層 model.add(Dense(units=1)) # 模型編譯 model.compile(optimizer='rmsprop', loss='mse') # 模型訓練 model.fit(X_train, y_train, epochs=20, batch_size=32) >>>Output exceeds the size limit. Open the full output data in a text editor Epoch 1/20 85/85 [==============================] - 30s 357ms/step - loss: 0.0666 Epoch 2/20 85/85 [==============================] - 32s 372ms/step - loss: 0.0124 Epoch 3/20 85/85 [==============================] - 31s 370ms/step - loss: 0.0080 Epoch 4/20 85/85 [==============================] - 34s 399ms/step - loss: 0.0063 Epoch 5/20 85/85 [==============================] - 45s 527ms/step - loss: 0.0055 Epoch 6/20 85/85 [==============================] - 45s 532ms/step - loss: 0.0048 Epoch 7/20 85/85 [==============================] - 46s 543ms/step - loss: 0.0039 Epoch 8/20 85/85 [==============================] - 45s 530ms/step - loss: 0.0039 Epoch 9/20 85/85 [==============================] - 47s 551ms/step - loss: 0.0034 Epoch 10/20 85/85 [==============================] - 49s 582ms/step - loss: 0.0032 Epoch 11/20 85/85 [==============================] - 49s 571ms/step - loss: 0.0029 Epoch 12/20 85/85 [==============================] - 48s 563ms/step - loss: 0.0026 Epoch 13/20 ... Epoch 19/20 85/85 [==============================] - 39s 460ms/step - loss: 0.0019 Epoch 20/20 85/85 [==============================] - 30s 351ms/step - loss: 0.0018 <tensorflow.python.keras.callbacks.History at 0x13f4a7ac8>

構建數據集，進行預測

dataset_total = pd.concat((dataset['High'][:"2016"], dataset['High']["2017":]), axis=0) dataset_total.shape >>>(3020,) dataset_total >>>Date 2006-01-03 82.55 2006-01-04 82.50 2006-01-05 82.90 2006-01-06 85.03 2006-01-09 84.25... 2017-12-22 153.00 2017-12-26 153.86 2017-12-27 153.18 2017-12-28 154.12 2017-12-29 154.72 Name: High, Length: 3020, dtype: float64 inputs = dataset_total[len(train_set):].values #測試集 inputs = inputs.reshape(-1, 1) ###轉換成一列 inputs.shape >>>(251, 1) inputs_scaled = sc.fit_transform(inputs) #歸一化

構建測試集

dataset_total = pd.concat((dataset['High'][:"2016"], dataset['High']["2017":]), axis=0) # 獲取輸入數據 inputs = dataset_total[len(dataset_total) - len(test_set) - 60:].values inputs >>>Output exceeds the size limit. Open the full output data in a text editor array([157.43, 157.7 , 158.49, 156.95, 154.97, 154.22, 155.53, 155.89,151. , 152.45, 152.9 , 151.15, 151.52, 151.16, 152.94, 154.06,154.44, 154.33, 153.91, 153.34, 153.74, 153.64, 156.11, 155.93,155.56, 161.16, 161.34, 161.86, 159.15, 159.55, 159.93, 160.72,163. , 163. , 162.38, 163.19, 164.66, 164.41, 163.8 , 162.2 ,160.29, 161.15, 160.79, 165.18, 166. , 166.72, 166.79, 169.95,169.89, 169.85, 169.11, 167.26, 168.25, 167.94, 168.23, 167.49,167.98, 167.74, 166.99, 166.7 , 167.87, 169.87, 169.39, 169.92,169.8 , 168.09, 167.76, 168.01, 168.48, 168.18, 168.59, 167.45,170.64, 171.25, 176. , 179.25, 178.88, 179.2 , 177.07, 175.58,175.7 , 174.97, 176.34, 175.98, 178.62, 177.5 , 177.8 , 178.87,179.9 , 180.13, 181.92, 182.79, 181.57, 180.79, 181.34, 182.5 ,181.49, 181.25, 180.63, 182.55, 181.88, 181.32, 180.99, 181.29,180.95, 179.25, 179.49, 178.06, 176.82, 176.28, 179. , 176.79,176.18, 176.23, 175.06, 175.67, 175.5 , 174.16, 175. , 174.49,174.59, 174.95, 174.87, 174.96, 176.33, 173.47, 172.93, 172.56,171.23, 171.2 , 171.36, 171.3 , 171.69, 162.64, 162.4 , 162.11,161.57, 162.04, 161.1 , 160.48, 160.59, 160.42, 159.49, 159.45,159.14, 155.78, 153.47, 153.1 , 152.37, 151.15, 151.26, 151.83,153.99, 154.14, 151.33, 152.46, 152.84, 153.68, 152.76, 153.73,153. , 152.67, 152.8 , 152.87, 153.2 , 152.93, 152.89, 154.2 ,152.82, 154.26, 157.2 , 155.48, 154.94, 154.69, 155.42, 155.86,155.39, 155.79, 154.99, 154.68, 155.75, 155.34, 155.55, 155.74,154.5 , 156.02, 155.89, 153.83, 153.49, 153.89, 153.65, 154.24,154.19, 154.62, 153.88, 154.29, 150.25, 148.83, 147.87, 147.04, ...153.47, 151.82, 151.51, 151.79, 151.8 , 150.89, 149. , 148.97,148.7 , 149.65, 150.15, 151.95, 152.45, 152.39, 152.2 , 152.49,152.93, 153.61, 154.41, 155.02, 156.8 , 156.74, 156.22, 154.45,155.03, 155.89, 157.85, 156.73, 155.11, 153.8 , 154.18, 154.17,153.89, 153.46, 153. , 153.86, 153.18, 154.12, 154.72]) # 歸一化 inputs = inputs.reshape(-1, 1) inputs = sc.transform(inputs) inputs.shape >>>(311, 1) # 準備測試集X_test,進行股價預測 X_test = [] for i in range(60, 311):X_test.append(inputs[i-60:i, 0])X_test = np.array(X_test) # numpy 類型轉換X_test.shape >>>(251, 60) ###構造數據集 X_test = np.reshape(X_test, (X_test.shape[0], X_test.shape[1], 1)) X_test.shape >>>(251, 60, 1) ###預測 predict_test = model.predict(X_test) # 預測 predict_test.shape >>>(251, 1) ###反歸一化 predict_stock_price = sc.inverse_transform(predict_test) predict_stock_price >>>([[168.57092],[168.69382],[169.65242],[170.35161],[170.84596],[171.02237],[170.4601 ],[169.75737],[169.38156],[169.46404],[169.58615],[169.80162],[169.56326],[170.3292 ],[171.34633],[173.63313],[176.39796],[178.01892],[178.7997 ],[178.39485],[177.41373],[176.70485],[176.20074],[176.44489],[176.7224 ], ...[154.5936 ],[154.30226],[154.3495 ],[154.2771 ],[154.49394]], dtype=float32)# 繪制測試結果和預測結果plot_predictions(test_set, predict_stock_price)

基于GRU

model_gru = Sequential() model_gru.add(GRU(50, return_sequences=True, input_shape=(X_train.shape[1], 1), activation='tanh')) model_gru.add(Dropout(0.2)) model_gru.add(GRU(50, activation='tanh')) model_gru.add(Dropout(0.2)) model_gru.add(Dense(1)) ###模型編譯 model_gru.compile(optimizer=SGD(lr=0.01, decay=1e-7, momentum=0.9), loss='mse' ###模型的訓練 model_gru.fit(X_train, y_train, epochs=20, batch_size=32) >>>Output exceeds the size limit. Open the full output data in a text editor Epoch 1/20 85/85 [==============================] - 18s 209ms/step - loss: 0.0351 Epoch 2/20 85/85 [==============================] - 18s 217ms/step - loss: 0.0032 Epoch 3/20 85/85 [==============================] - 19s 224ms/step - loss: 0.0026 Epoch 4/20 85/85 [==============================] - 19s 225ms/step - loss: 0.0023 Epoch 5/20 85/85 [==============================] - 19s 221ms/step - loss: 0.0021 Epoch 6/20 85/85 [==============================] - 19s 225ms/step - loss: 0.0019 Epoch 7/20 85/85 [==============================] - 19s 226ms/step - loss: 0.0018 Epoch 8/20 85/85 [==============================] - 19s 223ms/step - loss: 0.0017 Epoch 9/20 85/85 [==============================] - 19s 229ms/step - loss: 0.0016 Epoch 10/20 85/85 [==============================] - 19s 227ms/step - loss: 0.0015 Epoch 11/20 85/85 [==============================] - 19s 226ms/step - loss: 0.0016 Epoch 12/20 85/85 [==============================] - 15s 177ms/step - loss: 0.0014 Epoch 13/20 ... Epoch 19/20 85/85 [==============================] - 11s 135ms/step - loss: 0.0013 Epoch 20/20 85/85 [==============================] - 12s 139ms/step - loss: 0.0012 <tensorflow.python.keras.callbacks.History at 0x140e4feb8> # 準備測試集X_test,進行股價預測 X_test = [] for i in range(60, 311):X_test.append(inputs[i-60:i, 0])X_test = np.array(X_test) # numpy 類型轉換X_test = np.reshape(X_test, (X_test.shape[0], X_test.shape[1], 1)) #################### GRU_predicted = model_gru.predict(X_test) GRU_predicted_stock_price = sc.inverse_transform(GRU_predicted) # 可視化 plot_predictions(test_set, GRU_predicted_stock_price)

總結

以上是生活随笔為你收集整理的lstm单变量预测的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。