@@ -64,7 +64,9 @@ def createdataset(data):
6464
6565def myload ():
6666 filename = 'dump2.txt'
67- f = open ('data/' + filename , 'rb' )
67+ #f = open('./data/' + filename, 'rb')
68+ abspath = 'C:/Users/wwwa8/Documents/GitHub/Machine-Learning/序列预测/PCA去趋势化/dev2/data/'
69+ f = open (abspath + filename , 'rb' )
6870 data = pickle .load (f )
6971 f .close ()
7072 # print (data) # 路段数 * 每个路段的信息（df的数据结构）
@@ -102,14 +104,17 @@ def use_pca(dataset):
102104 time_step = 2
103105 data = myload () # data是一个list，里面是df格式的数据
104106 # transfer
105- dataset = createdataset (data ) # dataset 的格式是 （路段 * 每一天 * 一天内的数据）20 * 7 * 480
107+ dataset = createdataset (data ) # dataset 的格式是 （路段 * 每一天 * 一天内的数据）20 * 14 * 480
106108 dataset_main , dataset_rest = use_pca (dataset ) # dataset_main = 路段 * 每一个路段里面的主成分 ; dataset_rest = 路段 * 每一个路段里面的偏差
107109 days , dnum = dataset_main [0 ].shape
108- y_main = dataset_main [0 ][time_step :days , :]
109- trainX , trainY = split_dataset (dataset_rest )
110- train_x_raw = trainX [0 ]
111- train_y_raw = np .reshape (trainY [0 ], (days - time_step , 480 ))
112110
111+ y_main = dataset_main [0 ][time_step :days , :] # y的主成分[2~14] shape 12 * 2 * 480
112+ #将偏差数据 分割 成输入和输出
113+ trainX , trainY = split_dataset (dataset_rest ) #trainX 的shape 20 * 12 * 2 *480 ； trainY 的shape 20 * 12 * 1 *480
114+
115+ train_x_raw = trainX [0 ] #取了第一条路段来进行预测 train_x_raw的sahpe ：12 * 2 *480
116+ train_y_raw = np .reshape (trainY [0 ], (days - time_step , 480 )) # train_y_raw的shape 12 * 480
117+ # 归一化
113118 x_max = train_x_raw .max ()
114119 x_min = train_x_raw .min ()
115120 y_max = train_y_raw .max ()
@@ -118,6 +123,13 @@ def use_pca(dataset):
118123 train_x = (train_x_raw - x_min ) / (x_max - x_min )
119124 train_y = (train_y_raw - y_min ) / (y_max - y_min )
120125
126+ # test
127+ test_num = 3
128+ test_x = train_x [0 :test_num ]
129+ test_y = train_y [0 :test_num ]
130+
131+
132+ # 放入lstm训练
121133 # lstm的hyper-parameter
122134
123135 hidden_size = 400
@@ -126,7 +138,7 @@ def use_pca(dataset):
126138 dropout_keep_rate = 1
127139
128140 # 根据输入数据来决定，train_num训练集大小,input_size输入维度
129- train_num , time_step_size , input_size = train_x .shape
141+ train_num , time_step_size , input_size = train_x .shape # sahpe ：12 * 2 *480
130142 # output_size输出的结点个数
131143 _ , output_size = train_y .shape
132144
@@ -160,19 +172,19 @@ def use_pca(dataset):
160172 mae_result = []
161173 rmse_result = []
162174
163-
175+ # 获得训练的指标
164176 def get_metrics (y , pred_y ):
165177 mre = np .mean (np .abs (y - pred_y ) / y )
166178 mae = np .mean (np .abs (y - pred_y ))
167179 rmse = np .sqrt (np .mean (np .square (y - pred_y )))
168180 return mre , mae , rmse
169181
170182
171- def print_to_console (i , train_y_pred ):
172- train_y_pred_real = train_y_pred * (y_max - y_min ) + y_min
173- train_y_real = train_y * (y_max - y_min ) + y_min
174- plt .plot (range (dnum ), train_y_real [0 ], 'b-' )
175- plt .plot (range (dnum ), train_y_pred_real [0 ], 'r-' )
183+ def print_to_console (i , train_y , train_y_pred ):
184+ train_y_pred_real = train_y_pred * (y_max - y_min ) + y_min #反归一化
185+ train_y_real = train_y * (y_max - y_min ) + y_min #train_y 是真实的y值，堆train_y 进行反归一化
186+ plt .plot (range (dnum ), train_y_real [0 ], 'b-' ) #实际用蓝色
187+ plt .plot (range (dnum ), train_y_pred_real [0 ], 'r-' ) #预测用红色
176188 plt .show ()
177189 train_mre , train_mae , train_rmse = get_metrics (train_y_real , train_y_pred_real )
178190 print ('epoch %d train %.4f %.2f %.2f' % (i , train_mre , train_mae , train_rmse ))
@@ -184,7 +196,13 @@ def print_to_console(i, train_y_pred):
184196 if i % 50 == 0 :
185197 feed_dict = {x_input : train_x , y_real : train_y , keep_prob : 1.0 , batch_size : train_num }
186198 train_y_pred = sess .run (y_pred , feed_dict = feed_dict )
187- print_to_console (i , train_y_pred )
199+ print_to_console (i ,train_y , train_y_pred )
200+ if i % 50 == 0 :
201+ print ("test : " )
202+ feed_dict = {x_input : test_x , y_real : test_y , keep_prob : 1.0 , batch_size : test_num }
203+ test_y_pred = sess .run (y_pred , feed_dict = feed_dict )
204+ print_to_console (i , test_y ,test_y_pred )
205+ print ("--- test end ---" )
188206
189207 y_main = dataset_main [0 ][time_step :days , :]
190208 y_pre_real = y_main + train_y_pred * (y_max - y_min ) + y_min
0 commit comments