Skip to content

Commit 01d5688

Browse files
lawlite19lawlite19
committed
📝 修改神经网络部分bug
1 parent 7fc5684 commit 01d5688

4 files changed

Lines changed: 312 additions & 306 deletions

File tree

‎LinearRegression/LinearRegression.py‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -96,8 +96,8 @@ def plotJ(J_history,num_iters):
9696
# 测试linearRegression函数
9797
def testLinearRegression():
9898
mu,sigma,theta = linearRegression(0.01,400)
99-
print u"\n计算的theta值为：\n",theta
100-
print u"\n预测结果为：%f"%predict(mu, sigma, theta)
99+
#print u"\n计算的theta值为：\n",theta
100+
#print u"\n预测结果为：%f"%predict(mu, sigma, theta)
101101

102102
# 测试学习效果（预测）
103103
def predict(mu,sigma,theta):

‎NeuralNetwok/NeuralNetwork.py‎

Lines changed: 15 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -43,7 +43,7 @@ def neuralNetwork(input_layer_size,hidden_layer_size,out_put_layer):
4343
initial_nn_params = np.vstack((initial_Theta1.reshape(-1,1),initial_Theta2.reshape(-1,1))) #展开theta
4444
#np.savetxt("testTheta.csv",initial_nn_params,delimiter=",")
4545
start = time.time()
46-
result = optimize.fmin_cg(nnCostFunction, initial_nn_params, fprime=nnGradient, args=(input_layer_size,hidden_layer_size,out_put_layer,X,y,Lambda))
46+
result = optimize.fmin_cg(nnCostFunction, initial_nn_params, fprime=nnGradient, args=(input_layer_size,hidden_layer_size,out_put_layer,X,y,Lambda), maxiter=100)
4747
print (u'执行时间：',time.time()-start)
4848
print (result)
4949
'''可视化 Theta1'''
@@ -124,14 +124,16 @@ def nnCostFunction(nn_params,input_layer_size,hidden_layer_size,num_labels,X,y,L
124124
h = sigmoid(z3)
125125
'''代价'''
126126
J = -(np.dot(np.transpose(class_y.reshape(-1,1)),np.log(h.reshape(-1,1)))+np.dot(np.transpose(1-class_y.reshape(-1,1)),np.log(1-h.reshape(-1,1)))-Lambda*term/2)/m
127-
127+
#temp1 = (h.reshape(-1,1)-class_y.reshape(-1,1))
128+
#temp2 = (temp1**2).sum()
129+
#J = 1/(2*m)*temp2
128130
return np.ravel(J)
129131

130132
# 梯度
131133
def nnGradient(nn_params,input_layer_size,hidden_layer_size,num_labels,X,y,Lambda):
132134
length = nn_params.shape[0]
133-
Theta1 = nn_params[0:hidden_layer_size*(input_layer_size+1)].reshape(hidden_layer_size,input_layer_size+1)
134-
Theta2 = nn_params[hidden_layer_size*(input_layer_size+1):length].reshape(num_labels,hidden_layer_size+1)
135+
Theta1 = nn_params[0:hidden_layer_size*(input_layer_size+1)].reshape(hidden_layer_size,input_layer_size+1).copy() # 这里使用copy函数，否则下面修改Theta的值，nn_params也会一起修改
136+
Theta2 = nn_params[hidden_layer_size*(input_layer_size+1):length].reshape(num_labels,hidden_layer_size+1).copy()
135137
m = X.shape[0]
136138
class_y = np.zeros((m,num_labels)) # 数据的y对应0-9，需要映射为0/1的关系
137139
# 映射y
@@ -146,9 +148,8 @@ def nnGradient(nn_params,input_layer_size,hidden_layer_size,num_labels,X,y,Lambd
146148

147149
Theta1_grad = np.zeros((Theta1.shape)) #第一层到第二层的权重
148150
Theta2_grad = np.zeros((Theta2.shape)) #第二层到第三层的权重
149-
150-
Theta1[:,0] = 0;
151-
Theta2[:,0] = 0;
151+
152+
152153
'''正向传播，每次需要补上一列1的偏置bias'''
153154
a1 = np.hstack((np.ones((m,1)),X))
154155
z2 = np.dot(a1,np.transpose(Theta1))
@@ -157,15 +158,19 @@ def nnGradient(nn_params,input_layer_size,hidden_layer_size,num_labels,X,y,Lambd
157158
z3 = np.dot(a2,np.transpose(Theta2))
158159
h = sigmoid(z3)
159160

161+
160162
'''反向传播，delta为误差，'''
161163
delta3 = np.zeros((m,num_labels))
162164
delta2 = np.zeros((m,hidden_layer_size))
163165
for i in range(m):
164-
delta3[i,:] = h[i,:]-class_y[i,:]
166+
#delta3[i,:] = (h[i,:]-class_y[i,:])*sigmoidGradient(z3[i,:]) # 均方误差的误差率
167+
delta3[i,:] = h[i,:]-class_y[i,:] # 交叉熵误差率
165168
Theta2_grad = Theta2_grad+np.dot(np.transpose(delta3[i,:].reshape(1,-1)),a2[i,:].reshape(1,-1))
166169
delta2[i,:] = np.dot(delta3[i,:].reshape(1,-1),Theta2_x)*sigmoidGradient(z2[i,:])
167170
Theta1_grad = Theta1_grad+np.dot(np.transpose(delta2[i,:].reshape(1,-1)),a1[i,:].reshape(1,-1))
168171

172+
Theta1[:,0] = 0
173+
Theta2[:,0] = 0
169174
'''梯度'''
170175
grad = (np.vstack((Theta1_grad.reshape(-1,1),Theta2_grad.reshape(-1,1)))+Lambda*np.vstack((Theta1.reshape(-1,1),Theta2.reshape(-1,1))))/m
171176
return np.ravel(grad)
@@ -223,6 +228,7 @@ def checkGradient(Lambda = 0):
223228
step[i]=0
224229
# 显示两列比较
225230
res = np.hstack((num_grad.reshape(-1,1),grad.reshape(-1,1)))
231+
print("检查梯度的结果，第一列为数值法计算得到的，第二列为BP得到的:")
226232
print (res)
227233

228234
# 初始化调试的theta权重
@@ -256,5 +262,5 @@ def predict(Theta1,Theta2,X):
256262
return p
257263

258264
if __name__ == "__main__":
259-
#checkGradient()
265+
checkGradient()
260266
neuralNetwork(400, 25, 10)

0 commit comments

Comments
 (0)