写了几百年的神经网络终于写出来了。
主要遇到的几个坑点:
1.数据集里的y label必须在(0,1)间,如果你用了sigmoid作为激活函数
2.对于反向传播 S(L)=y-a(L) sigmoid函数 S(l)=w(l).S(l+1)*a(l)*(1-a(l)) sigmoid的导数
其实这里的坑点是,后面都是两两相乘,一开始用异或,输出都是1,就没注意到差别
delta(l)=S(l+1).a(l)
#-*- encoding:utf-8 -*- import sys import numpy as np import os import matplotlib.pyplot as plt import random from mlxtend.evaluate import plot_decision_regions from matplotlib import animation reload(sys) sys.setdefaultencoding('utf-8') class bpnn: def __init__(self,a,b,c,d): np.random.seed(1) self.l0_w=np.random.random((a,b)) #输出层到隐含层的权值 self.l1_w=np.random.random((b,c)) #隐含1层到隐含2层的权值 self.l2_w=np.random.random((c,d)) #隐含2层到输出层的权值 def train(self,X,Y,iterators,test_percent,lr): prev_loss=0 flag=0 for i in xrange(iterators+1): j=random.randint(0,len(X)-1) x=X[j] #取一行训练数据 y=Y[j] l0=x l1=sigmoid(np.dot(x,self.l0_w)) #得到第一层输出 l2=sigmoid(np.dot(l1,self.l1_w)) #得到第二层输出 l3=sigmoid(np.dot(l2,self.l2_w)) l3_error=l3-y loss=np.sqrt(np.sum(np.square(l3_error)))*100 l2_delta=l2.T.dot(l3_error) self.l2_w-=lr*l2_delta l3_error=l3_error.transpose() l2_error=np.multiply(self.l2_w.dot(l3_error),sigmoid(l2,True).T) #不能省略求导,已经是输出了,所以直接相乘 l1_error=np.multiply(self.l1_w.dot(l2_error),sigmoid(l1,True).T) #不能省略求导 l1_delta=l1.T.dot(l2_error.T) # l0_delta=l0.T.dot(l1_error.T) self.l1_w-=lr*l1_delta self.l0_w-=lr*l0_delta if(flag%10000==0): if(flag!=0): plt.plot([flag-10000,flag],[prev_loss,loss],'r--') plt.scatter(flag,loss,s=40,c='red') print "tranning:"+str(flag)+" times"+",total loss="+str(loss) prev_loss=loss flag=flag+1 def predict(self,x): x=AutoNorm(x) l1=sigmoid(np.dot(x,self.l0_w)) l2=sigmoid(np.dot(l1,self.l1_w)) l3=sigmoid(np.dot(l2,self.l2_w)) return l3 def AutoNorm(Z): Zmax,Zmin = Z.max(), Z.min() Z = (Z - Zmin)/(Zmax - Zmin) return Z def sigmoid(x,deriv=False): y=1.0 / (1 + np.exp(-x)) if(deriv==True): return np.multiply(x,1-x) return y def load_dataset(dataset,len_x,len_y): dataset=open(dataset,'r') line=dataset.readline() X=list() Y=list() while line: l=line.strip('\n') l=l.split(' ') l = [ float( l ) for l in l if l ] X.append(l[0:len_x]) Y.append(l[len_x:len_x+len_y]) line=dataset.readline() return (X,Y) if __name__=="__main__": degree1=4 degree2=4 newnn=bpnn(4,degree1,degree2,2) (X,Y)=load_dataset('E:\\ml\\iris.txt',4,2) X=np.matrix(X) Y=np.matrix(Y) X=AutoNorm(X) y=list() for i in Y: y.append(i.tolist()[0][0]) plt.figure(2) plt.xlabel("epoch times") plt.ylabel("loss rate") plt.title("total loss rate graph") #ax1=plt.subplot(211) #ax2=plt.subplot(212) newnn.train(X,Y,60000,0.3,0.035) #plt.sca(ax1) #plot_decision_regions(X,y,newnn,legend=1) #plt.title("Decision Boundary for hidden layer size %d"%degree) plt.savefig('E:\\ml\\iris.jpg') print newnn.predict(np.matrix([6.7,3.1,4.4,1.4])) plt.show()训练损失
