DBN Python预测交通流

import numpy as np from K_Means import kMeans from DBN import * import csv import tensorflow as tf from switch_case import switch training_set_k = [] reader = csv.reader(open('pemstrain.csv','r')) for row in reader: training_set_k.append(row) #print(training_set_k[1][0]) training_set_1 = [[training_set_k[0][0],training_set_k[1][0],training_set_k[2][0],training_set_k[3][0]], [training_set_k[0][1],training_set_k[1][1],training_set_k[2][1],training_set_k[3][1]], [training_set_k[0][2],training_set_k[1][2],training_set_k[2][2],training_set_k[3][2]], [training_set_k[0][3],training_set_k[1][3],training_set_k[2][3],training_set_k[3][3]], [training_set_k[0][4],training_set_k[1][4],training_set_k[2][4],training_set_k[3][4]], [training_set_k[0][5],training_set_k[1][5],training_set_k[2][5],training_set_k[3][5]], [training_set_k[0][6],training_set_k[1][6],training_set_k[2][6],training_set_k[3][6]], [training_set_k[0][7],training_set_k[1][7],training_set_k[2][7],training_set_k[3][7]], [training_set_k[0][8],training_set_k[1][8],training_set_k[2][8],training_set_k[3][8]], [training_set_k[0][9],training_set_k[1][9],training_set_k[2][9],training_set_k[3][9]]] training_set_1n = np.array(training_set_1,dtype="float64") reader = csv.reader(open('pemstrainlabel.csv','r')) training_label_k = [] for row in reader: training_label_k.append(row) # print(training_label_k[0][0]) training_label_1 = [[training_label_k[0][0],training_label_k[1][0],training_label_k[2][0],training_label_k[3][0]], [training_label_k[0][1], training_label_k[1][1], training_label_k[2][1], training_label_k[3][1]], [training_label_k[0][2], training_label_k[1][2], training_label_k[2][2], training_label_k[3][2]], [training_label_k[0][3], training_label_k[1][3], training_label_k[2][3], training_label_k[3][3]], [training_label_k[0][4], training_label_k[1][4], training_label_k[2][4], training_label_k[3][4]], [training_label_k[0][5], training_label_k[1][5], training_label_k[2][5], training_label_k[3][5]], [training_label_k[0][6], training_label_k[1][6], training_label_k[2][6], training_label_k[3][6]], [training_label_k[0][7], training_label_k[1][7], training_label_k[2][7], training_label_k[3][7]], [training_label_k[0][8], training_label_k[1][8], training_label_k[2][8], training_label_k[3][8]], [training_label_k[0][9], training_label_k[1][9], training_label_k[2][9], training_label_k[3][9]]] training_label_1n = np.array(training_label_1,dtype="float64") #print(training_label_1n.dtype) testing_set_k = [] reader = csv.reader(open('pemstest.csv','r')) for row in reader: testing_set_k.append(row) #print(testing_set_k[1][0]) testing_set_1 = [[testing_set_k[0][0],testing_set_k[1][0],testing_set_k[2][0],testing_set_k[3][0]], [testing_set_k[0][1],testing_set_k[1][1],testing_set_k[2][1],testing_set_k[3][1]]] testing_set_1n = np.array(testing_set_1,dtype="float64") reader = csv.reader(open('pemstrainlabel.csv','r')) testing_label_k = [] for row in reader: testing_label_k.append(row) # print(testing_label_k[0][0]) testing_label_1 = [[testing_label_k[0][0],testing_label_k[1][0],testing_label_k[2][0],testing_label_k[3][0]], [testing_label_k[0][1], testing_label_k[1][1], testing_label_k[2][1], testing_label_k[3][1]]] testing_label_1n = np.array(testing_label_1,dtype="float64") #print(training_set_1) dbn = DBN(input=training_set_1n,label=training_label_1n,n_ins=4 , hidden_layer_sizes=[128,128,128], n_outs= 4) dbn.pretrain(lr=0.3,k=1,epochs=40) dbn.finetune(0.1,40) # print(dbn.get_last_layer().W) center,ClusterAssiment = kMeans(dbn.rbm_layers[-1].W,k=2) # print(ClusterAssiment) # print(counter) # print(dbn.log_layer.W.T) training_set_1t1 = [] training_set_1t2 = [] training_label_1t1 = [] training_label_1t2 = [] testing_set_1t1 = [] testing_set_1t2 = [] counter1=[] counter2 = [] for i in range(len(dbn.log_layer.W.T)): if ClusterAssiment[i,0] == 0.0: training_set_1t1.append(training_set_1n[:,i]) training_label_1t1.append(training_label_1n[:,i]) testing_set_1t1.append(testing_set_1n[:,i]) counter1.append(i) else: training_set_1t2.append(training_set_1n[:,i]) training_label_1t2.append(training_label_1n[:,i]) testing_set_1t2.append(testing_set_1n[:,i]) counter2.append(i) training_set_1t1_n = np.array(training_set_1t1) training_set_1t2_n = np.array(training_set_1t2) training_label_1t1_n = np.array(training_label_1t1) training_label_1t2_n = np.array(training_label_1t2) # print(training_label_1t1_n.T) # print(training_label_1t1_n[0,:]) # print(training_set_1n) # print(len(training_set_1t1_n)) # print(training_set_1t2_n.T) # print(training_label_1t1_n.T) if len(training_set_1t1_n)!=0 and len(training_set_1t2_n)!=0: dbn1 = DBN(input=training_set_1t1_n.T,label=training_label_1t1_n.T,n_ins=len(training_set_1t1_n),hidden_layer_sizes=[16],n_outs=len(training_set_1t1_n)) dbn1.pretrain(0.3,1,2) dbn1.finetune(0.1,2) dbn2 = DBN(input=training_set_1t2_n.T,label=training_label_1t2_n.T,n_ins=len(training_set_1t2_n),hidden_layer_sizes=[16],n_outs=len(training_set_1t2_n)) dbn2.pretrain(0.3,1,2) dbn2.finetune(0.1,2) elif len(training_set_1t1_n)==0: dbn2 = DBN(input=training_set_1t2_n.T, label=training_label_1t2_n.T, n_ins=len(training_set_1t2_n), hidden_layer_sizes=[128,128,128], n_outs=len(training_set_1t2_n)) dbn2.pretrain(0.3, 2) dbn2.finetune() else: dbn1 = DBN(input=training_set_1t1_n.T, label=training_label_1t1_n.T, n_ins=len(training_set_1t1_n), hidden_layer_sizes=[128,128,128], n_outs=len(training_set_1t1_n)) dbn1.pretrain(0.3, 2) dbn1.finetune() print(dbn.log_layer.W) def share_train(epchos,training_x1,training_x2,initial_shared_weights_1,initial_shared_weights_2,training_label_1,training_label_2,Y1_shared_weights,Y2_shared_weights,Y3_shared_weights,Y4_shared_weights): d_y_1 = [] d_y_2 = [] y1_given_x1 = softmax(np.dot(training_x1, initial_shared_weights_1)) y2_given_x1 = softmax(np.dot(training_x2, initial_shared_weights_2)) # y_given_x1 = softmax(np.dot(training_x,initial_shared_weights)) # print(initial_shared_weights_1.shape) # print(Y1_shared_weights.shape) # print(training_x1.shape) # print(training_x2.shape) # print(y1_given_x1.shape) # print(training_label_1t1_n.T.shape) # print(y1_given_x1.shape) for epcho in range(epchos): # p_y_given_x1 = softmax(np.dot(y_given_x1,Y1_shared_weights)) for i in range(len(counter1)): if counter1[i] == 0: p_y1_given_x_1 = softmax(np.dot(y1_given_x1, Y1_shared_weights)) d_y1 = training_label_1[i,:].T.reshape(10,1)-p_y1_given_x_1 # d_y_1.append(d_y1) # print(p_y1_given_x_1.shape) Y1_shared_weights += 0.1 * np.dot(y1_given_x1.T,d_y1) elif counter1[i] == 1: p_y2_given_x_1 = softmax(np.dot(y1_given_x1, Y2_shared_weights)) d_y2 = training_label_1[i,:].T.reshape(10,1)-p_y2_given_x_1 # d_y_1.append(d_y2) # print(p_y2_given_x_1.shape) Y2_shared_weights += 0.1 * np.dot(y1_given_x1.T,d_y2) elif counter1[i] == 2: p_y3_given_x_1 = softmax(np.dot(y1_given_x1, Y3_shared_weights)) d_y3 = training_label_1[i,:].T.reshape(10,1)-p_y3_given_x_1 # d_y_1.append(d_y3) # print(p_y3_given_x_1.shape) Y3_shared_weights += 0.1 * np.dot(y1_given_x1.T,d_y3) elif counter1[i] == 3: p_y4_given_x_1 = softmax(np