我正在尝试从头开始构建神经网络，以在最小的数据库上进行图像识别，但是代码不仅产生了错误的答案，而且似乎还遥不可及。
我试图在中间步骤中查看权重和偏差，但是由于无法知道应该是什么权重和偏差，所以我不知道如何解决。
import tensorflow as tf
import matplotlib.pyplot as plt
(xtrain,ytrain),(xtest,ytest)=tf.keras.datasets.mnist.load_data()
import numpy as np
import random
flatlist1=[x.flatten() for x in xtrain]
flatlist2=[x.flatten() for x in xtest]
traininp=[]
testinp=[]
for i in range(len(ytrain)):
traininp.append([flatlist1[i],ytrain[i]])
for i in range(len(ytest)):
testinp.append([flatlist1[i],ytest[i]])
def sigmoid(x):
return 1.0/(1.0+np.exp(-x))
def sigmoid_der(x):
return sigmoid(x)*(1-sigmoid(x))
def relu(x):
return max(0,x)
def relu_der(x):
if x>0:
return 1
else:
return 0
class NeuralNetwork(object):
def __init__(self,traininp,layers):
#Add a list layer provideing no of nodes in each layer
#add training input as array of pixel grayscale values and output in a list
self.num_layers = len(layers)
self.layers = layers.insert(0,len(traininp[0][0]))
self.weights=[]
self.biases = [((np.random.randn(y, 1)).flatten()).tolist() for y in layers[1:]]
for y in range(1,len(layers[1:])+1):
weight=[]
for x in range(layers[y]):
weight.append(((np.random.randn(layers[y-1],1)).flatten()).tolist() )
self.weights.append(weight)
self.nweights=[]
self.nbiases=[]
for i in range(len(self.weights)):
w=[]
for j in self.weights[i]:
w.append([0]*len(j))
self.nweights.append(w)
for i in self.biases:
self.nbiases.append([0]*len(i))
x=0
for i in range(10):
for k in range(20):
predictions=self.feedforward(traininp[x][0],layers)
dbiases,dweights=self.Backprop(traininp[x][1],predictions)
self.nbiases=self.addbiases(self.nbiases,dbiases)
self.nweights=self.addweights(self.nweights,dweights)
x+=1
self.updateb(self.biases,self.nbiases,0.1,20)
self.updatew(self.weights,self.nweights,0.1,20)
for i in range(190,210):
predictions=self.feedforward(traininp[i][0],layers)
print(predictions[-1],predictions[-1].index(max(predictions[-1])),traininp[i][1])
def calculate(self,inp,weights,bias):
wsum=0.0
for x,y in zip(weights,inp):
wsum+=x*y
return (wsum+bias)
def feedforward(self,traininp,layers):
firstz=[self.calculate(traininp,self.weights[0][i],self.biases[0][i]) for i in range(layers[1])]
firstlayer=[sigmoid(i) for i in firstz]
outputlayers=[]
outputlayers.append((traininp).tolist())
self.zs=[]
self.zs.append(firstz)
outputlayers.append(firstlayer)
for x in range(1,len(layers[1:])):
z=[]
k=[]
for i in range(layers[x+1]):
z.append(self.calculate(outputlayers[x-1],self.weights[x][i],self.biases[x][i]))
k.append(sigmoid(z[i]))
self.zs.append(z)
outputlayers.append(k)
return(outputlayers)
def Backprop(self,traininp,predictions):
dweights=[]
dbiases=[]
dnodes=[]
dweights=self.nweights.copy()
dbiases=self.nbiases.copy()
y=[0]*len(predictions[-1])
y[traininp]=1
for i in predictions:
dnodes.append([0]*len(i))
for i in range(len(y)):
dbiases[-1][i]=sigmoid_der(self.zs[-1][i])*self.cost(predictions[-1][i],y[i])
for j in range(len(predictions[-2])):
dweights[-1][i][j]=predictions[-2][j]*sigmoid_der(self.zs[-1][i])*self.cost(predictions[-1][i],y[i])
dweights[-1][i][j]/=j
for j in range(len(self.weights[-1])):
for k in range(len(self.weights[-1][j])):
dnodes[-1][i]=self.weights[-1][j][k]*sigmoid_der(self.zs[-1][i])*self.cost(predictions[-1][i],y[i])
dnodes[-1][i]/=k
for k in range(2,self.num_layers+1):
for i in range(len(predictions[-k])):
dbiases[-k][i]=sigmoid_der(self.zs[-k][i])*self.cost(predictions[-k][i],dnodes[-k][i])
for j in range(len(self.weights[-k])):
for l in range(len(self.weights[-k][j])):
dnodes[-k][i]=self.weights[-k][j][l]*sigmoid_der(self.zs[-k][i])*self.cost(predictions[-k][i],dnodes[-k][i])
dnodes[-k][i]/=l
for j in range(len(predictions[-k-1])):
dweights[-k][i][j]=predictions[-k-1][j]*sigmoid_der(self.zs[-k][i])*self.cost(predictions[-k][i],dnodes[-k][i])
dweights[-k][i][j]/=j
return (dbiases,dweights)
def cost(self,output,y):
return (2*(output-y))
def addbiases(self,list1,list2):
listo=[]
for i,j in zip(list1,list2):
listt=[]
for k,l in zip(i,j):
listt.append(k+l)
listo.append(listt)
return listo
def addweights(self,list1,list2):
listo=[]
for i,j in zip(list1,list2):
lists=[]
for k,l in zip(i,j):
listt=[]
for p,q in zip(k,l):
listt.append(p+q)
lists.append(listt)
listo.append(lists)
return listo
def updatew(self,list1,list2,eta,bsize):
listo=[]
for i,j in zip(list1,list2):
lists=[]
for k,l in zip(i,j):
listt=[]
for p,q in zip(k,l):
listt.append(p-(eta/bsize)*q)
lists.append(listt)
listo.append(lists)
return listo
def updateb(self,list1,list2,eta,bsize):
listo=[]
for i,j in zip(list1,list2):
listt=[]
for k,l in zip(i,j):
listt.append(k-(eta/bsize)*l)
listo.append(listt)
return listo
n1=NeuralNetwork(traininp[0:210],[5,5,10])
我至少希望它像神经网络一样工作，即使它的准确性不高或工作速度很慢。