bp.rar_BP_bp神经_python 神经网络_神经网络python

__author__ = 'sjs' import numpy as np from datetime import datetime import os from PIL import Image def new_str(): return datetime.now().strftime('%b-%d-%y %H:%M:%S') def logistic(x): return 1 / (1 + np.exp(-x)) def logistic_derivative(x): return x*(1-x) def tanh(x): return np.tanh(x) def tanh_derivative(x): return 1-x**2 class NetLayer: ''' 网络层封装 管理当前网络层的神经元列表 ''' #神经网络层初始化 def __init__(self, len_node, in_count, default='sigmoid',mu=0,sigma=0.5,bias=True): ''' :param len_node: 当前层的神经元数 :param in_count: 当前层的输入数 :bias为偏置项 ''' #如果是sigmoid if default=='sigmoid': self.nonlinear = (logistic, logistic_derivative) #如果是tanh elif default=='tanh': self.nonlinear = (tanh, tanh_derivative) self.bias = bias #如果bias为真，则输入层数加1 if bias: in_count += 1# bias #权重矩阵初始化 self.weights = np.reshape(np.random.normal(mu, sigma, len_node*in_count), (len_node, in_count)) #残差初始化 self.deltas_item = np.zeros(len_node) # 记录下一层的引用，方便递归操作 self.next_layer = None #输入输出初始化 self.input = np.array([]) self.output = np.array([]) #递归计算输出值 def calc_output(self, x): if self.bias: self.input = np.hstack((x, np.ones((len(x),1)))) else: self.input = x self.output = self.nonlinear[0](np.dot(self.input, self.weights.T)) if self.next_layer is not None: return self.next_layer.calc_output(self.output) return self.output #递归更新权重 def update_weight(self, learning_rate, target): ''' 更新当前网络层及之后层次的权重 使用了递归来操作，所以要求外面调用时必须从网络层的第一层（输入层的下一层）来调用 :param learning_rate: 学习率 :param target: 输出值 ''' self.deltas_item = self.nonlinear[1](self.output) if self.next_layer is None: self.deltas_item *= (target - self.output ) else: self.deltas_item *= self.next_layer.update_weight(learning_rate, target) #计算出weight应当变化的量，w_add的大小为当前层weight的大小，即i*j，这个值是len(target)个变量作用的结果 w_add = np.dot(self.deltas_item.T, self.input) w_add *= learning_rate self.weights +=w_add weights = self.weights if not self.bias else self.weights[:, :-1] return np.dot(self.deltas_item, weights) class NeuralNetWork: #初始化，构造输入规模的神经网络 def __init__(self, layers): self.initLayers=[] self.layers = [] self.construct_network(layers) self.kind=layers[-1] pass def construct_network(self, layers): last_layer = None for i, layer in enumerate(layers): if i == 0: continue cur_layer = NetLayer(layer, layers[i-1]) self.layers.append(cur_layer) if last_layer is not None: last_layer.next_layer = cur_layer last_layer = cur_layer def fit(self, x_train, y_train,learning_rate=0.01, epochs=500,mode='shuffle'): ''' 训练网络, 默认按顺序来训练 方法 1：按训练数据顺序来训练 方法 2: 随机选择测试 :param x_train: 输入数据 :param y_train: 输出数据 :param learning_rate: 学习率 :param epochs:权重更新次数 ''' y_trainformat=format_y(y_train,self.kind) if mode=='shuffle': indices = np.arange(len(y_train)) np.random.shuffle(indices) length = 100 accper=0.0 cishu=0 while accper < 0.95: for index in range(len(indices)//length): indexes = indices[index*length: (index+1)*length] self.layers[0].calc_output(x_train[indexes]) self.layers[0].update_weight(learning_rate, y_trainformat[indexes]) accper=self.show_test(x_train,y_train) print(new_str(),"epochs", cishu) print('准确率',self.show_test(x_train,y_train)) cishu +=1 if cishu > epochs: break def show_test(self, x_test, y_test): predict = self.predict(x_test) predicts = [] for p in predict: predicts.append(unformat_y(p)) predicts = np.array(predicts) ok_cnt = sum([1 for i in range(len(predicts)) if predicts[i] == y_test[i]]) return float(ok_cnt)/len(predicts) def predict(self, x): return self.layers[0].calc_output(x) def format_y(data,kind): y_tmp = np.zeros((len(data), kind)) for i in range(len(data)): y_tmp[i, data[i]] = 1 return y_tmp def unformat_y(data): #找到最大的那一项 m = max(data) for j, val in enumerate(data): if val == m: return j #出现意外返回-1 return -1 def getimdata(imfile): #读取图片 im=Image.open(imfile) return list(map((lambda x:x/255),im.getdata())) def getDigitData(patharg): X=list() Y=list() filelist=os.listdir(patharg) for i in filelist: #flag为这张图片代表几 flag=i.split('_')[0] Y.append(int(flag)) resx=getimdata(patharg+i) X.append(resx) return np.array(X),np.array(Y) #start X,Y=getDigitData(r'smallhandscriptdata/') Xtest,Ytest=getDigitData(r'test/') bp=NeuralNetWork([28*28,400,300,10]) bp.fit(X,Y) print('###########################################') res=bp.predict(Xtest) for i in range(len(res)): print('实际结果:',Ytest[i],' 预测结果:',unformat_y(res[i]))