second

集中式/center1.py

@@ -4,19 +4,25 @@
 @Date   ：2020/11/29  10:10
 @Desc   ：
 """
+import datetime
+import os
+import time
 
 import torch
+from scipy.interpolate import make_interp_spline
+from statsmodels.tsa.seasonal import seasonal_decompose
 from torch import nn
 import numpy
 import pandas as pd
-from pandas import read_csv
 from sklearn.preprocessing import MinMaxScaler
 from functools import reduce
 from log import logger
+from tensorboardX import SummaryWriter
+import matplotlib.pyplot as plt
 
 look_back = 4
 EPOCH = 10
-SIZE = 50
+SIZE = 150
 dataset = None
 Nums = 0
 
@@ -54,13 +60,27 @@ def evluation(real, pred):
 #         return data_X
 
 # 初始化数据
-dataset = numpy.zeros((30, 50), dtype=float)
-dataset[1] = numpy.random.poisson(lam=20, size=SIZE)
-dataset[2] = numpy.random.poisson(lam=10, size=SIZE)
-dataset[4] = numpy.random.poisson(lam=30, size=SIZE)
+dataset = numpy.zeros((30, SIZE), dtype=float)
+d1 = numpy.random.poisson(lam=20.0, size=SIZE)
+dataset[1] = d1
+dataset[2] = numpy.random.poisson(lam=20.0, size=SIZE)
+dataset[4] = numpy.random.poisson(lam=20.0, size=SIZE)
+
+# 返回包含三个部分 trend（趋势部分），seasonal（季节性部分）和residual (残留部分)
+def decompose(timeseries):
+    decomposition = seasonal_decompose(timeseries, freq=3)
+    trend = decomposition.trend
+    seasonal = decomposition.seasonal
+    residual = decomposition.resid
+    return trend, seasonal, residual
+
 def get_data(node_id, time_series):
     time_series = numpy.array(time_series)
-    dataset[node_id - 1] = time_series
+    time_series = pd.Series(numpy.log(time_series + 1).reshape(-1), index=pd.DatetimeIndex(range(len(time_series))))
+    trend, seasonal, residual = decompose(time_series)
+    ts = trend + seasonal
+    ts = ts.fillna(numpy.nanmean(numpy.array(ts)))
+    dataset[node_id - 1] = ts.values
 
 # 判断参与的节点有哪些
 def join_train(dataset=dataset):
@@ -99,7 +119,9 @@ def forward(self, x):
         x = x.view(s, b, -1)
         return x
 
+
 class TrainPredict(object):
+
     def __init__(self):
         self.var_x, self.var_y, self.model = None, None, None
         self.Node = []
@@ -112,7 +134,7 @@ def transfrom(self):
             data_X, data_Y = load_data1(data.reshape(-1, 1))
             dfs_x.append(data_X)
             dfs_y.append(data_Y.reshape(-1, 1))
-        dataset = dataset = numpy.zeros((30, 50), dtype=float)
+        dataset = dataset = numpy.zeros((30, SIZE), dtype=float)
         data_X = reduce(lambda left, right: numpy.concatenate((left, right), axis=1), dfs_x)
         data_Y = reduce(lambda left, right: numpy.concatenate((left, right), axis=1), dfs_y)
         train_X = data_X
@@ -121,27 +143,52 @@ def transfrom(self):
         train_Y = train_Y.reshape(-1, 1, train_Y.shape[1])
         self.var_x = torch.from_numpy(train_X)
         self.var_y = torch.from_numpy(train_Y)
-        self.model = LSTM(train_X.shape[-1], 64, Nums, 1)
+        return train_X.shape[-1]
         # print(self.model)
-        print(dataset)
+        # print(dataset)
 
     def train(self):
-        self.transfrom()
+        input_shape = self.transfrom()
+
+        self.model = LSTM(input_shape, Nums*look_back*4, Nums, 1)
         loss_fun = nn.MSELoss()
         optimizer = torch.optim.Adam(self.model.parameters(), lr=0.05)
+
+        writer = SummaryWriter("logs")
+        writer.add_graph(self.model, input_to_model=torch.zeros(1, 1, Nums*look_back), verbose=True)
+        loss_plot = [1]
+        time_plot = [0]
+        first = 0
         for epoch in range(1, EPOCH+1):
             logger.info("Epoch: %d......" % epoch)
-            for t in range(50):
+            if not first:
+                first = float(time.time())
+                writer.add_scalar("train loss", 1.0, 0)
+            for t in range(25):
                 # 前向传播
                 out = self.model(self.var_x)
                 loss = loss_fun(out, self.var_y)
                 # 反向传播
                 optimizer.zero_grad()
                 loss.backward()
                 optimizer.step()
+            # tensorboard --logdir=logs
+            writer.add_scalar("train loss", loss.item(), float(time.time()*1000) - first*1000)
+            time_plot.append(float(time.time()) - first)
+            loss_plot.append(loss.item())
+
+        x_smooth = numpy.linspace(min(time_plot), max(time_plot), 300)
+        y_smooth = make_interp_spline(numpy.array(time_plot), numpy.array(loss_plot))(x_smooth)
+        plt.plot(x_smooth, y_smooth)
+        plt.savefig('plot.jpg')
+        plt.show()
+
+        return time_plot, loss_plot
+
 
     # 预测结果，传入参数需要得到预测值的个数
-    def center_predict(self, node_id, time=50):
+    def center_predict(self, node_id, time=150):
+        torch.save(self.model, "model.pkl")
         # self.train()
         # 返回二维数组，索引减一代表节点号对应的预测结果值
         predicts = []
@@ -150,8 +197,10 @@ def center_predict(self, node_id, time=50):
         # time是需要的结果个数
         for _ in range(time):
             predict = self.model(cur_test_x)
-            predicts.append(predict.data.numpy().reshape(-1).tolist())
+            # for j in range(len(predict[0][0])):
+            #     predict[0][0][j] = cur_test_x[0][0][j: j + look_back].mean() * 0.8 + predict[0][0][j] * 0.2
 
+            predicts.append(predict.data.numpy().reshape(-1).tolist())
             # 更改下一时刻的特征
             cur = []
             for i in range(1, Nums*look_back+1):
@@ -160,14 +209,18 @@ def center_predict(self, node_id, time=50):
                 else:
                     cur.append(predict[0][0][i//look_back-1].data)
             cur_test_x = torch.from_numpy(numpy.array(cur)).view(-1, 1, look_back*Nums)
+
         predicts = scaler.inverse_transform(numpy.array(predicts).reshape(Nums, -1))
-        print(predicts.shape)
+
         result = numpy.zeros((30, time))
         for i, node in enumerate(self.Node):
             result[node] = predicts[i]
         return result[node_id]
 
 tp = TrainPredict()
-tp.train()
-result = tp.center_predict(1, 50)
-print(result)
+time_plot, loss_plot = tp.train()
+
+result = tp.center_predict(1, 150)
+r = numpy.concatenate((d1, result), axis=0)
+
+