#-*- coding: utf-8 -*- ''' - - - - - -- - - - - - - - - - - - - - - - - - - - - - - Name - - CNN - Convolution Neural Network For Photo Recognizing Goal - - Recognize Handing Writting Word Photo Detail：Total 5 layers neural network * Convolution layer * Pooling layer * Input layer layer of BP * Hiden layer of BP * Output layer of BP Author: Stephen Lee Program: PYTHON Date: 2017.9.20 - - - - - -- - - - - - - - - - - - - - - - - - - - - - - ''' import numpy as np import matplotlib.pyplot as plt class CNN(): conv1 = [] w_conv1 = [] thre_conv1 = [] step_conv1 = 0 size_pooling1 = 0 num_bp1 = 0 num_bp2 = 0 num_bp3 = 0 thre_bp1 = [] thre_bp2 = [] wkj = np.mat([]) vji = np.mat([]) rate_weight = 0 rate_thre = 0 def __init__(self,conv1_get,size_p1,bp_num1,bp_num2,bp_num3,rate_w=0.2,rate_t=0.2): ''' :param conv1_get: [a,c,d]，size, number, step of convolution kernel :param size_p1: pooling size :param bp_num1: units number of flatten layer :param bp_num2: units number of hidden layer :param bp_num3: units number of output layer :param rate_w: rate of weight learning :param rate_t: rate of threshold learning ''' self.num_bp1 = bp_num1 self.num_bp2 = bp_num2 self.num_bp3 = bp_num3 self.conv1 = conv1_get[:2] self.step_conv1 = conv1_get[2] self.size_pooling1 = size_p1 self.rate_weight = rate_w self.rate_thre = rate_t self.w_conv1 = [np.mat(-1*np.random.rand(self.conv1[0],self.conv1[0])+0.5) for i in range(self.conv1[1])] self.wkj = np.mat(-1 * np.random.rand(self.num_bp3, self.num_bp2) + 0.5) self.vji = np.mat(-1*np.random.rand(self.num_bp2, self.num_bp1)+0.5) self.thre_conv1 = -2*np.random.rand(self.conv1[1])+1 self.thre_bp2 = -2*np.random.rand(self.num_bp2)+1 self.thre_bp3 = -2*np.random.rand(self.num_bp3)+1 def save_model(self,save_path): #将模型保存 import pickle model_dic = {'num_bp1':self.num_bp1, 'num_bp2':self.num_bp2, 'num_bp3':self.num_bp3, 'conv1':self.conv1, 'step_conv1':self.step_conv1, 'size_pooling1':self.size_pooling1, 'rate_weight':self.rate_weight, 'rate_thre':self.rate_thre, 'w_conv1':self.w_conv1, 'wkj':self.wkj, 'vji':self.vji, 'thre_conv1':self.thre_conv1, 'thre_bp2':self.thre_bp2, 'thre_bp3':self.thre_bp3} with open(save_path, 'wb') as f: pickle.dump(model_dic, f) print('模型已经保存： %s'% save_path) def paste_model(self,save_path): #实例方法， #虽然这么写一点也不简洁。。。。 #卸载这个里面的话，只是用于修改已经存在的模型，要根据读取的数据返回实例的模型，再写一个吧 import pickle with open(save_path, 'rb') as f: model_dic = pickle.load(f) self.num_bp1 = model_dic.get('num_bp1') self.num_bp2 = model_dic.get('num_bp2') self.num_bp3 = model_dic.get('num_bp3') self.conv1 = model_dic.get('conv1') self.step_conv1 = model_dic.get('step_conv1') self.size_pooling1 = model_dic.get('size_pooling1') self.rate_weight = model_dic.get('rate_weight') self.rate_thre = model_dic.get('rate_thre') self.w_conv1 = model_dic.get('w_conv1') self.wkj = model_dic.get('wkj') self.vji = model_dic.get('vji') self.thre_conv1 = model_dic.get('thre_conv1') self.thre_bp2 = model_dic.get('thre_bp2') self.thre_bp3 = model_dic.get('thre_bp3') print('已经成功读取模型') @classmethod def ReadModel(cls,model_path): #类方法，读取保存的模型，返回一个实例。 import pickle with open(model_path, 'rb') as f: model_dic = pickle.load(f) conv_get= model_dic.get('conv1') conv_get.append(model_dic.get('step_conv1')) size_p1 = model_dic.get('size_pooling1') bp1 = model_dic.get('num_bp1') bp2 = model_dic.get('num_bp2') bp3 = model_dic.get('num_bp3') r_w = model_dic.get('rate_weight') r_t = model_dic.get('rate_thre') #创建实例 conv_ins = CNN(conv_get,size_p1,bp1,bp2,bp3,r_w,r_t) #修改实例的参数 conv_ins.w_conv1 = model_dic.get('w_conv1') conv_ins.wkj = model_dic.get('wkj') conv_ins.vji = model_dic.get('vji') conv_ins.thre_conv1 = model_dic.get('thre_conv1') conv_ins.thre_bp2 = model_dic.get('thre_bp2') conv_ins.thre_bp3 = model_dic.get('thre_bp3') return conv_ins def sig(self,x): return 1 / (1 + np.exp(-1*x)) def do_round(self,x): return round(x, 3) #卷积 def Convolute(self,data,convs,w_convs,thre_convs,conv_step): size_conv = convs[0] num_conv =convs[1] size_data = np.shape(data)[0] #得到原图像滑动的小图，data_focus data_focus = [] for i_focus in range(0, size_data - size_conv + 1, conv_step): for j_focus in range(0, size_data - size_conv + 1, conv_step): focus = data[i_focus:i_focus + size_conv, j_focus:j_focus + size_conv] data_focus.append(focus) #计算所有卷积核得到的特征图，每个特征图以矩阵形式，存储为一个列表data_featuremap data_featuremap = [] Size_FeatureMap = int((size_data - size_conv) / conv_step + 1) for i_map in range(num_conv): featuremap = [] for i_focus in range(len(data_focus)): net_focus = np.sum(np.multiply(data_focus[i_focus], w_convs[i_map])) - thre_convs[i_map] featuremap.append(self.sig(net_focus)) featuremap = np.asmatrix(featuremap).reshape(Size_FeatureMap, Size_FeatureMap) data_featuremap.append(featuremap) #将data_focus中的focus展开为一维 focus1_list = [] for each_focus in data_focus: focus1_list.extend(self.Expand_Mat(each_focus)) focus_list = np.asarray(focus1_list) return focus_list,data_featuremap # 池化 def Pooling(self,featuremaps,size_pooling): size_map = len(featuremaps[0]) size_pooled = int(size_map/size_pooling) featuremap_pooled = [] for i_map in range(len(featuremaps)): map = featuremaps[i_map] map_pooled = [] for i_focus in range(0,size_map,size_pooling): for j_focus in range(0, size_map, size_pooling): focus = map[i_focus:i_focus + size_pooling, j_focus:j_focus + size_pooling] #平均池化 map_pooled.append(np.average(focus)) #最大池化 #map_pooled.append(np.max(focus)) map_pooled = np.asmatrix(map_pooled).reshape(size_pooled,size_pooled) featuremap_pooled.append(map_pooled) return featuremap_pooled def Expand(self,datas): #将三元的数据展开为1为的list data_expanded = [] for i in range(len(datas)): shapes = np.shape(datas[i]) data_listed = datas[i].reshape(1,shapes[0]*shapes[1]) data_listed = data_listed.getA().tolist()[0] data_expanded.extend(data_listed) #连接所有数据 data_expanded = np.asarray(data_expanded) return data_expanded def Expand_Mat(self,data_mat): #用来展开矩阵为一维的list data_mat = np.asarray(data_mat) shapes = np.shape(data_mat) data_expanded = data_mat.reshape(1,shapes[0]*shapes[1]) return data_expanded def Getpd_From_Pool(self,out_map,pd_pool,num_map,size_map,size_pooling): ''' 误差反传，从pooled到前一个map, 例如将池化层6*6的误差矩阵扩大为12*12的误差矩阵 pd_pool: 是采样层的误差，list形式。。。。要改要改 out_map: 前面特征图的输出，数量*size*size的列表形式 return: pd_all:前面层所有的特征图的pd， num*size_map*size_map的列表形式 ''' pd_all = [] i_pool = 0 for i_map in range(num_map): pd_conv1 = np.ones((size_map, size_map)) for i in range(0, size_map, size_pooling): for j in range(0, size_map, size_pooling): pd_conv1[i:i + size_pooling, j:j + size_pooling] = pd_pool[i_pool] i_pool = i_pool + 1 pd_conv2 = np.multiply(pd_conv1,np.multiply(out_map[i_map],(1-out_map[i_map]))) pd_all.append(pd_conv2) return pd_all def trian(self,patterns,datas_train, datas_teach, n_repeat, error_accuracy,draw_e = bool): print('----------------------Start Training-------------------------') print(' - - Shape: Train_Data ',np.shape(datas_train)) print(' - - Shape: Teach_Data ',np.shape(datas_teach)) rp = 0 all_mse = [] mse = 10000 while rp < n_repeat and mse >= error_accuracy: alle = 0 print('-------------进行第%d次学习--------------'%rp) for p in range(len(datas_train)): #print('------------学习第%d个图像--------------'%p) data_train = np.asmatrix(datas_train[p]) data_teach = np.asarray(datas_teach[p]) data_focus1,data_conved1 = self.Convolute(data_train,self.conv1,self.w_conv1, self.thre_conv1,conv_step=self.step_conv1) data_pooled1 = self.Pooling(data_conved1,self.size_pooling1) shape_featuremap1 = np.shape(data_conved1) ''' print(' -----original shape ', np.shape(data_train)) print(' ---- after convolution ',np.shape(data_conv1)) print(' -----after pooling ',np.shape(data_pooled1)) ''' data_bp_input = self.Expand(data_pooled1) # 计算第一层输入输出 bp_out1 = data_bp_input # 计算第二层输入输出 bp_net_j = np.dot(bp_out1,self.vji.T) - self.thre_bp2 bp_out2 = self.sig(bp_net_j) # 计算第三层输入输出 bp_net_k = np.dot(bp_out2 ,self.wkj.T) - self.thre_bp3 bp_out3 = self.sig(bp_net_k) # 计算一般化误差 pd_k_all = np.multiply((data_teach - bp_out3), np.multiply(bp_out3, (1 - bp_out3))) pd_j_all = np.multiply(np.dot(pd_k_all,self.wkj), np.multiply(bp_out2, (1 - bp_out2))) pd_i_all = np.dot(pd_j_all,self.vji) pd_conv1_pooled = pd_i_all / (self.size_pooling1*self.size_pooling1) pd_conv1_pooled = pd_conv1_pooled.T.getA().tolist() pd_conv1_all = self.Getpd_From_Pool(data_conved1,pd_conv1_pooled,shape_featuremap1[0], shape_featuremap1[1],self.size_pooling1) #卷积层1的权重和阈值修正，每个卷积核的权重需要修正 12*12(map) 次 #修正量为featuremap中点的偏导值 乘以 前一层图像focus， 整个权重模板一起更新 for k_conv in range(self.conv1[1]): pd_conv_list = self.Expand_Mat(pd_conv1_all[k_conv]) delta_w = self.rate_weight * np.dot(pd_conv_list,data_focus1) self.w_conv1[k_conv] = self.w_conv1[k_conv] + delta_w.reshape((self.conv1[0],self.conv1[0])) self.thre_conv1[k_conv] = self.thre_conv1[k_conv] - np.sum(pd_conv1_all[k_conv]) * self.rate_thre # 更新kj层的权重 self.wkj = self.wkj + pd_k_all.T * bp_out2 * self.rate_weight # 更新ji层的权重 self.vji = self.vji + pd_j_all.T * bp_out1 * self.rate_weight # 更新阈值 self.thre_bp3 = self.thre_bp3 - pd_k_all * self.rate_thre self.thre_bp2 = self.thre_bp2 - pd_j_all * self.rate_thre # 计算总误差 errors = np.sum(abs((data_teach - bp_out3))) alle = alle + errors #print(' ----Teach ',data_teach) #print(' ----BP_output ',bp_out3) rp = rp + 1 mse = alle/patterns all_mse.append(mse) def draw_error(): yplot = [error_accuracy for i in range(int(n_repeat * 1.2))] plt.plot(all_mse, '+-') plt.plot(yplot, 'r--') plt.xlabel('Learning Times') plt.ylabel('All_mse') plt.grid(True, alpha=0.5) plt.show() print('------------------Training Complished---------------------') print(' - - Training epoch: ', rp, ' - - Mse: %.6f' % mse) if draw_e: draw_error() return mse def produce(self,datas_test): #对验证和测试数据集进行输出 produce_out = [] print('-------------------Start Testing-------------------------') print(' - - Shape: Test_Data ',np.shape(datas_test)) for p in range(len(datas_test)): print('--------测试第%d个图像----------' % p) data_test = np.asmatrix(datas_test[p]) data_focus1, data_conved1 = self.Convolute(data_test, self.conv1, self.w_conv1, self.thre_conv1, conv_step=self.step_conv1) data_pooled1 = self.Pooling(data_conved1, self.size_pooling1) data_bp_input = self.Expand(data_pooled1) # 计算第一层输入输出 bp_out1 = data_bp_input # 计算第二层输入输出 bp_net_j = bp_out1 * self.vji.T - self.thre_bp2 bp_out2 = self.sig(bp_net_j) # 计算第三层输入输出 bp_net_k = bp_out2 * self.wkj.T - self.thre_bp3 bp_out3 = self.sig(bp_net_k) produce_out.extend(bp_out3.getA().tolist()) res = [list(map(self.do_round,each)) for each in produce_out] return np.asarray(res) def convolution(self,data): #返回卷积和池化后的数据,用于查看图像 data_test = np.asmatrix(data) data_focus1, data_conved1 = self.Convolute(data_test, self.conv1, self.w_conv1, self.thre_conv1, conv_step=self.step_conv1) data_pooled1 = self.Pooling(data_conved1, self.size_pooling1) return data_conved1,data_pooled1