# 導(dao)入所需(xu)庫(ku)
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
import torch.optim as optim
from torch.autograd import Variable

df_app = pd.read_csv('app-user-activiation-data.csv', index_col='Date', parse_dates=['Date']) #導入數(shu)據
# 按照2020年10月1日為界拆分數據集
Train = df_app[:'2020-09-30'].iloc[:,0:1].values #訓練集(ji)
Test = df_app['2020-10-01':].iloc[:,0:1].values #測試集
from sklearn.preprocessing import MinMaxScaler #導(dao)入歸一化縮放器
Scaler = MinMaxScaler(feature_range=(0,1)) #創建縮放器
Train = Scaler.fit_transform(Train) #擬合縮放(fang)器并對訓練(lian)集進(jin)行(xing)歸(gui)一化(hua)
# 對測試數據進行歸一化處(chu)理
Test = Scaler.transform(Test)

import matplotlib.pyplot as plt #導入matplotlib.pyplot
plt.style.use('fivethirtyeight') #設定(ding)繪圖風格
df_app["Activation"].plot(figsize=(12,4),legend=True) #繪制激(ji)活數
plt.title('App Activation Count') #圖(tu)題
plt.show() #繪圖

# 創建一(yi)個函數(shu)，將數(shu)據集(ji)轉化為時間序列格式
def sliding_windows(data, seq_length):
    x = []
    y = []
    for i in range(len(data)-seq_length):
        _x = data[i:(i+seq_length)]
        _y = data[i+seq_length]
        x.append(_x)
        y.append(_y)
    return np.array(x),np.array(y)
# 設(she)定窗口大小
seq_length = 4
x_train, y_train = sliding_windows(Train, seq_length)
# 使用滑動窗口(kou)為測試(shi)數據創建特征和標簽(qian)
x_test, y_test = sliding_windows(Test, seq_length)
# 將數據轉化為torch張(zhang)量
testX = Variable(torch.Tensor(np.array(x_test)))
testY = Variable(torch.Tensor(np.array(y_test)))

# 設置模(mo)型參數
input_size = 1
hidden_size = 64
num_layers = 1
output_size = 1

class RNN(nn.Module):
    def __init__(self, input_size, hidden_size, num_layers, output_size):
        super(RNN, self).__init__()
        
        self.hidden_size = hidden_size
        
        # RNN層(ceng)
        self.rnn = nn.RNN(input_size, hidden_size, num_layers, batch_first=True)
        
        # 全連接(jie)層(ceng)，用于輸(shu)出
        self.fc = nn.Linear(hidden_size, output_size)
    def forward(self, x):
        # 初始化隱(yin)狀(zhuang)態
        h0 = Variable(torch.zeros(num_layers, x.size(0), self.hidden_size))
        
        # 前向傳播RNN
        out, _ = self.rnn(x, h0)
        
        # 解碼RNN的最后一(yi)個隱藏層的輸出
        out = self.fc(out[:, -1, :])
        
        return out

# 創建模型
rnn = RNN(input_size, hidden_size, num_layers, output_size)
# 定(ding)義損失函數和優(you)化(hua)器
criterion = torch.nn.MSELoss()    # 均方誤差
optimizer = torch.optim.Adam(rnn.parameters(), lr=0.01)   # Adam優化器(qi)
# 將數據轉化為(wei)torch張量
trainX = Variable(torch.Tensor(np.array(x_train)))
trainY = Variable(torch.Tensor(np.array(y_train)))
# 訓(xun)練(lian)模(mo)型(xing)
num_epochs = 100
for epoch in range(num_epochs):
    outputs = rnn(trainX)
    optimizer.zero_grad()
    
    # 計算損失
    loss = criterion(outputs, trainY)
    loss.backward()
    
    optimizer.step()
    if epoch % 10 == 0:
        print("Epoch: %d, loss: %1.5f" % (epoch, loss.item()))

# 使用訓練好的模型進(jin)行預測
rnn."> 設置模型為評估模式
test_outputs = rnn(testX)
# 將預測(ce)結果逆歸一化(hua)
test_outputs = test_outputs.data.numpy()
test_outputs = Scaler.inverse_transform(test_outputs) # 逆歸一化
# 真實(shi)測(ce)試標簽逆歸(gui)一化
y_test_actual = Scaler.inverse_transform(y_test)
# 輸(shu)出預測和真實結果
for i in range(len(y_test)):
    print(f"Date: {df_app['2020-10-01':].index[i+seq_length]}, Actual Activation: {y_test_actual[i][0]}, Predicted Activation: {test_outputs[i][0]}")

Date: 2020-10-05 00:00:00, Actual Activation: 923.0, Predicted Activation: 885.0889282226562
Date: 2020-10-06 00:00:00, Actual Activation: 919.0000000000001, Predicted Activation: 893.8333129882812
Date: 2020-10-07 00:00:00, Actual Activation: 915.0, Predicted Activation: 898.3600463867188
Date: 2020-10-08 00:00:00, Actual Activation: 910.0000000000001, Predicted Activation: 896.8088989257812
Date: 2020-10-09 00:00:00, Actual Activation: 905.0000000000001, Predicted Activation: 895.2501220703125
Date: 2020-10-10 00:00:00, Actual Activation: 901.0, Predicted Activation: 891.626953125
Date: 2020-10-11 00:00:00, Actual Activation: 913.0000000000001, Predicted Activation: 888.2245483398438
Date: 2020-10-12 00:00:00, Actual Activation: 900.0, Predicted Activation: 890.2631225585938
Date: 2020-10-13 00:00:00, Actual Activation: 888.0000000000001, Predicted Activation: 885.708984375
Date: 2020-10-14 00:00:00, Actual Activation: 883.0, Predicted Activation: 880.345458984375
Date: 2020-10-15 00:00:00, Actual Activation: 861.0, Predicted Activation: 878.3155517578125
Date: 2020-10-16 00:00:00, Actual Activation: 844.0, Predicted Activation: 865.739013671875
Date: 2020-10-17 00:00:00, Actual Activation: 837.0, Predicted Activation: 853.5955810546875
Date: 2020-10-18 00:00:00, Actual Activation: 841.0, Predicted Activation: 845.1204223632812
Date: 2020-10-19 00:00:00, Actual Activation: 821.0, Predicted Activation: 838.4027709960938
Date: 2020-10-20 00:00:00, Actual Activation: 843.0, Predicted Activation: 826.8444213867188
Date: 2020-10-21 00:00:00, Actual Activation: 857.0, Predicted Activation: 830.8392333984375
Date: 2020-10-22 00:00:00, Actual Activation: 861.0, Predicted Activation: 837.3353881835938
Date: 2020-10-23 00:00:00, Actual Activation: 858.0, Predicted Activation: 838.6307983398438
Date: 2020-10-24 00:00:00, Actual Activation: 832.0, Predicted Activation: 845.0518798828125
Date: 2020-10-25 00:00:00, Actual Activation: 811.0, Predicted Activation: 839.138427734375
Date: 2020-10-26 00:00:00, Actual Activation: 807.0, Predicted Activation: 828.6616821289062
Date: 2020-10-27 00:00:00, Actual Activation: 803.0, Predicted Activation: 820.3754272460938
Date: 2020-10-28 00:00:00, Actual Activation: 821.0, Predicted Activation: 809.3426513671875
Date: 2020-10-29 00:00:00, Actual Activation: 838.0, Predicted Activation: 809.6136474609375
...Date: 2021-01-22 00:00:00, Actual Activation: 925.0000000000001, Predicted Activation: 897.9528198242188
Date: 2021-01-23 00:00:00, Actual Activation: 926.0, Predicted Activation: 898.5119018554688
Date: 2021-01-24 00:00:00, Actual Activation: 920.0, Predicted Activation: 899.7281494140625
Date: 2021-01-25 00:00:00, Actual Activation: 932.0, Predicted Activation: 899.1580200195312

# 定義(yi)繪圖函(han)數(shu)
def plot_predictions(test,predicted):
    plt.plot(test, color='red',label='Real Count') #真(zhen)值(zhi)
    plt.plot(predicted, color='blue',label='Predicted Count') #預測值(zhi)
    plt.title('Flower App Activation Prediction') #圖題
    plt.xlabel('Time') #X軸時間
    plt.ylabel('Flower App Activation Count') #Y軸激活(huo)數
    plt.legend() #圖例
    plt.show() #繪(hui)圖
plot_predictions(y_test_actual,test_outputs) #繪圖

import math #導入數學函數
from sklearn.metrics import mean_squared_error

def return_rmse(test,predicted): #定義均方損失(shi)函數
    rmse = math.sqrt(mean_squared_error(test, predicted)) #均方損失(shi)
    print("MSE損失值 {}".format(rmse))
return_rmse(y_test_actual, test_outputs)

from sklearn.metrics import r2_score

r2 = r2_score(y_test_actual, test_outputs)
print(f"R2 Score: {r2}")