我正在尝试制作一个可以从其他程序导入的神经网络Python模块。我用kaggle上的某些数据来测试这个模块。然而，无论输入是什么，输出几乎总是相同的，误差在1e-7以内！！我已经对数据进行了归一化，并且尝试增加了更多的隐藏层和更多的隐藏节点。我还降低了学习率，并且在训练过程中使其进一步降低。我尝试添加偏置，但这产生了负面效果
这里是可导入的hidden_net.py模块:
class network: def __init__(self,layer_num,learning_rate=0.7,seed=None,logistic_coefficent=0.9): self.logistic_coefficent=logistic_coefficent self.learning_rate=learning_rate self.w0 = np.random.random((layer_num[0],layer_num[1])) self.w1 = np.random.random((layer_num[1],layer_num[2])) np.random.seed(seed) def sigmoid(self,x,reverse=False): if(reverse==True): return x*(1-x) return 1/(1+np.exp(-x)) def train(self,inps,outs,mod_learn_rate=0): if mod_learn_rate == 0: mod_learn_=self.learning_rate inps=np.array(inps) layer0 = inps layer1 = self.sigmoid(np.dot(layer0,self.w0)) layer2 = self.sigmoid(np.dot(layer1,self.w1)) layer2_error = outs - layer2 layer2_delta = layer2_error*self.sigmoid(layer2,reverse=True)#*mod_learn_rate layer1_error = layer2_delta.dot(self.w1.T) layer1_delta = layer1_error * self.sigmoid(layer1,reverse=True)#*mod_learn_rate self.w1 += layer2.T.dot(layer2_delta) self.w0 += layer1.T.dot(layer1_delta) return np.mean(abs(layer2_error)) def calcout(self,inp): inp=np.array(inp) layer0=inp layer1=self.sigmoid(np.dot(layer0,self.w0)) out=self.sigmoid(np.dot(layer1,self.w1)) return out以及导入该模块的脚本:
import randomfrom numpy import meanrandom.seed(50404)op=open('Mall_Customers_Mod.txt','r')full=op.read()op.close()full_lines=full.split('\n')training_lines=random.sample(full_lines,175)training_inputs=[]training_outputs=[]for j in training_lines: training_inputs.append([float(j.split(',')[0]),float(j.split(',')[1]),float(j.split(',')[2])]) training_outputs.append(float(j.split(',')[3]))testing_lines=random.sample(full_lines,175)testing_inputs=[]testing_outputs=[]for l in testing_lines: testing_inputs.append([float(l.split(',')[0]),float(l.split(',')[1]),float(j.split(',')[2])]) testing_outputs.append(float(l.split(',')[3]))nn=hidden_net.network([3,9,1],seed=50404,learning_rate=10)er=[]txt=''try: for i in range(10000): for l in range(len(training_inputs)): er.append(nn.train(training_inputs[l],training_outputs[l],10/(i+1))) if (i%1000==0 or i==1 or i==0 or i==2): print('epoch:{}\nerror:{}\nlearning_rate={}'.format(i,mean(er),10/(i+1))) txt=txt+'\nepoch:{}\nerror:{}'.format(i,mean(er)) er=[]except KeyboardInterrupt: passprint('done!')score=0error=[]tests=0for i in range(len(testing_inputs)): print('net output: ' +str(nn.calcout(testing_inputs[i]))) print('true output: '+str(testing_outputs[i])) error.append(abs(nn.calcout(testing_inputs[i]) - testing_outputs[i]))print('error: {}'.format(mean(error)))print('\n\nweights:{}'.format(nn.w0))归一化后的数据格式为
Sex Age Income Spending Score0,0.019230769230769232,0.0,0.38775510204081630,0.057692307692307696,0.0,0.81632653061224491,0.038461538461538464,0.00819672131147541,0.051020408163265311,0.09615384615384616,0.00819672131147541,0.77551020408163261,0.25,0.01639344262295082,0.3979591836734694我本以为输出会有所不同，但事实并非如此
net output: [0.49777196]true output: 0.3979591836734694net output: [0.49999907]true output: 0.8571428571428571net output: [0.49997918]true output: 0.7346938775510204net output: [0.49983941]true output: 0.30612244897959184net output: [0.49999985]这似乎是神经网络中一个相当常见的错误，原因有很多。我本想增加悬赏但我负担不起。尝试了各种方法很久了，希望你们中的某一位能解决这个问题！提前感谢，3NiGMa
回答：
你的学习率对于你的数据集大小来说太高了，试着将其提高到0.1或1，这样你的权重可以更快地变化
