Python神经网络不训练

Question

我有一个简单的神经网络，有2个输入神经元，3个隐藏神经元和1个输出神经元。隐藏层有偏差

我不习惯矩阵运算做前馈和反向传播。当我在一个简单的线性数据集上运行训练函数时，会出现错误，并且预测结果是错误的

import random
from math import exp,pow,tanh

def random_weight():
    return random.random()

def sigmoid(x):
    return 1.0 / (1.0 + exp(-x))

def sigmoid_drv(x):
    return sigmoid(x)*(1.0-sigmoid(x))

w11_I = random_weight()
w12_I = random_weight()
w21_I = random_weight()
w22_I = random_weight()
w31_I = random_weight()
w32_I = random_weight()
w11_II = random_weight()
w12_II = random_weight()
w13_II = random_weight()
b_I = 1

activation = sigmoid
activation_drv = sigmoid_drv

def predict(x1,x2):
    global w11_I,w12_I,w21_I,w22_I,w31_I,w32_I,w11_II,w12_II,w13_II,b_I
    a1_I = w11_I*x1 + w12_I*x2 + b_I
    z1_I = activation(a1_I)
    a2_I = w21_I*x1 + w22_I*x2 + b_I
    z2_I = activation(a2_I)
    a3_I = w31_I*x1 + w32_I*x2 + b_I
    z3_I = activation(a3_I)
    a1_II = w11_II*z1_I + w12_II*z2_I + w13_II*z3_I
    z1_II = activation(a1_II)
    return a1_I, z1_I, a2_I, z2_I, a3_I, z3_I, a1_II, z1_II

def train(x1,x2,y,alpha):
    global w11_I,w12_I,w21_I,w22_I,w31_I,w32_I,w11_II,w12_II,w13_II,b_I
    a1_I, z1_I, a2_I, z2_I, a3_I, z3_I, a1_II, z1_II = predict(x1,x2)
    error = 0.5 * pow(y-z1_II,2)
    delta = y-z1_II * activation_drv(a1_II)

    w11_II += delta * z1_I * alpha
    w12_II += delta * z2_I * alpha
    w13_II += delta * z3_I * alpha

    w11_I += delta * w11_II * activation_drv(a1_I) * x1 * alpha
    w12_I += delta * w11_II * activation_drv(a1_I) * x2 * alpha
    w21_I += delta * w12_II * activation_drv(a2_I) * x1 * alpha
    w22_I += delta * w12_II * activation_drv(a2_I) * x2 * alpha
    w31_I += delta * w13_II * activation_drv(a3_I) * x1 * alpha
    w32_I += delta * w13_II * activation_drv(a3_I) * x2 * alpha

    b_I += (delta * w11_II * activation_drv(a1_I) + delta * w12_II * activation_drv(a2_I) + delta * w13_II * activation_drv(a3_I)) * alpha
    return error

data = [
    [0,0,0],
    [0,1,1],
    [1,0,1],
    [1,1,1],
]

for i in range(0,10):
    err = 0
    dt = data[::]
    random.shuffle(dt)
    for j in dt:
        err += train(j[0],j[1],j[2],0.01)
    print(err)

print("-"*30)
for j in data:
    _, _, _, _, _, _, _, res = predict(j[0],j[1])
    print(j[0],",",j[1],"=",res)

例如，代码的结果是：

0.363894453262
0.366966815948
0.366406041572
0.369982058232
0.36988850637
0.375869833099
0.378106172616
0.380456639936
0.37901554717
0.383723920259
------------------------------
(0, ',', 0, '=', 0.8439871540493414)
(0, ',', 1, '=', 0.861714406183168)
(1, ',', 0, '=', 0.8515477541104413)
(1, ',', 1, '=', 0.8676931366534011)

------------更新------------------

我将代码更改为：

import random
from math import exp,pow

def random_weight():
    return random.random()

def sigmoid(x):
    return 1.0 / (1.0 + exp(-x))

def sigmoid_drv(x):
    return sigmoid(x)*(1.0-sigmoid(x))

w11_I = random_weight()
w12_I = random_weight()
w21_I = random_weight()
w22_I = random_weight()
w31_I = random_weight()
w32_I = random_weight()
w11_II = random_weight()
w12_II = random_weight()
w13_II = random_weight()
b_I = random_weight()

activation = sigmoid
activation_drv = sigmoid_drv

def predict(x1,x2):
    global w11_I,w12_I,w21_I,w22_I,w31_I,w32_I,w11_II,w12_II,w13_II,b_I
    a1_I = w11_I*x1 + w12_I*x2 + b_I
    z1_I = activation(a1_I)
    a2_I = w21_I*x1 + w22_I*x2 + b_I
    z2_I = activation(a2_I)
    a3_I = w31_I*x1 + w32_I*x2 + b_I
    z3_I = activation(a3_I)
    a1_II = w11_II*z1_I + w12_II*z2_I + w13_II*z3_I
    z1_II = activation(a1_II)
    return a1_I, z1_I, a2_I, z2_I, a3_I, z3_I, a1_II, z1_II

def train(x1,x2,y,alpha):
    global w11_I,w12_I,w21_I,w22_I,w31_I,w32_I,w11_II,w12_II,w13_II,b_I
    a1_I, z1_I, a2_I, z2_I, a3_I, z3_I, a1_II, z1_II = predict(x1,x2)
    error = 0.5 * pow(z1_II-y,2)
    delta = z1_II-y * activation_drv(a1_II)

    d_w11_II = delta * z1_I * alpha 
    d_w12_II = delta * z2_I * alpha
    d_w13_II = delta * z3_I * alpha

    d_w11_I = delta * w11_II * activation_drv(a1_I) * x1 * alpha
    d_w12_I = delta * w11_II * activation_drv(a1_I) * x2 * alpha
    d_w21_I = delta * w12_II * activation_drv(a2_I) * x1 * alpha
    d_w22_I = delta * w12_II * activation_drv(a2_I) * x2 * alpha
    d_w31_I = delta * w13_II * activation_drv(a3_I) * x1 * alpha
    d_w32_I = delta * w13_II * activation_drv(a3_I) * x2 * alpha
    d_b_I = (delta * w11_II * activation_drv(a1_I) + delta * w12_II * activation_drv(a2_I) + delta * w13_II * activation_drv(a3_I)) * alpha

    w11_II -= d_w11_II
    w12_II -= d_w12_II
    w13_II -= d_w13_II

    w11_I -= d_w11_I
    w12_I -= d_w12_I
    w21_I -= d_w21_I
    w22_I -= d_w22_I
    w31_I -= d_w31_I
    w32_I -= d_w32_I
    b_I -= d_b_I

    return error

data = [
    [0,0,0],
    [0,1,0],
    [1,0,0],
    [1,1,1],
]
for i in range(0,10):
    err = 0
    dt = data[::]
    random.shuffle(dt)
    for j in dt:
        err += train(j[0],j[1],j[2],0.01)
    print(err)

print("-"*30)
for j in data:
    _, _, _, _, _, _, _, res = predict(j[0],j[1])
    print(j[0],",",j[1],"=",res)

我现在用重量减去重量误差。网络误差减小。但预测仍然是错误的

上述代码的结果：

0.7793443881847488
0.7577581315356949
0.7432698222320477
0.7316129719356839
0.7160385688813552
0.6943522088277978
0.6862277294774705
0.6656984495700775
0.6584361784187711
0.6410006126876817
------------------------------
0 , 0 = 0.6049212721996029
0 , 1 = 0.6227402202339664
1 , 0 = 0.6139758543180651
1 , 1 = 0.6293581473456563

Answer 1

我发现了问题 sigmoid函数不适合此网络。我把它改为tanh，现在预测结果是正确的

最终代码：

import random
from math import exp,pow

class ANN:
    def random_weight(self):
        return random.random()

    def sigmoid(self,x):
        return 1.0 / (1.0 + exp(-x))

    def sigmoid_drv(self,x):
        return self.sigmoid(x)*(1.0-self.sigmoid(x))  

    def tanh(self, x):
        return (exp(x) - exp(-x)) / (exp(x) + exp(-x))

    def tanh_drv(self,x):
        return 1 - pow(self.tanh(x),2)

    def __init__(self):
        self.w11_I = self.random_weight()
        self.w12_I = self.random_weight()
        self.w21_I = self.random_weight()
        self.w22_I = self.random_weight()
        self.w31_I = self.random_weight()
        self.w32_I = self.random_weight()
        self.w11_II = self.random_weight()
        self.w12_II = self.random_weight()
        self.w13_II = self.random_weight()
        self.b_I = self.random_weight()
        self.activation = self.tanh
        self.activation_drv = self.tanh_drv

    def predict(self,x1,x2):
        a1_I = self.w11_I*x1 + self.w12_I*x2 + self.b_I
        z1_I = self.activation(a1_I)
        a2_I = self.w21_I*x1 + self.w22_I*x2 + self.b_I
        z2_I = self.activation(a2_I)
        a3_I = self.w31_I*x1 + self.w32_I*x2 + self.b_I
        z3_I = self.activation(a3_I)
        a1_II = self.w11_II*z1_I + self.w12_II*z2_I + self.w13_II*z3_I
        z1_II = self.activation(a1_II)
        return a1_I, z1_I, a2_I, z2_I, a3_I, z3_I, a1_II, z1_II

    def train(self,x1,x2,y,alpha):
        a1_I, z1_I, a2_I, z2_I, a3_I, z3_I, a1_II, z1_II = self.predict(x1,x2)
        error = 0.5 * pow(z1_II-y,2)
        delta = (z1_II-y) * self.activation_drv(a1_II)

        d_w11_II = delta * z1_I * alpha 
        d_w12_II = delta * z2_I * alpha
        d_w13_II = delta * z3_I * alpha

        d_w11_I = delta * self.w11_II * self.activation_drv(a1_I) * x1 * alpha
        d_w12_I = delta * self.w11_II * self.activation_drv(a1_I) * x2 * alpha
        d_w21_I = delta * self.w12_II * self.activation_drv(a2_I) * x1 * alpha
        d_w22_I = delta * self.w12_II * self.activation_drv(a2_I) * x2 * alpha
        d_w31_I = delta * self.w13_II * self.activation_drv(a3_I) * x1 * alpha
        d_w32_I = delta * self.w13_II * self.activation_drv(a3_I) * x2 * alpha
        d_b_I = (delta * self.w11_II * self.activation_drv(a1_I) + delta * self.w12_II * self.activation_drv(a2_I) + delta * self.w13_II * self.activation_drv(a3_I)) * alpha

        self.w11_II -= d_w11_II
        self.w12_II -= d_w12_II
        self.w13_II -= d_w13_II

        self.w11_I -= d_w11_I
        self.w12_I -= d_w12_I
        self.w21_I -= d_w21_I
        self.w22_I -= d_w22_I
        self.w31_I -= d_w31_I
        self.w32_I -= d_w32_I
        self.b_I -= d_b_I

        return error


model = ANN()

data = [
    [0,0,0],
    [0,1,0],
    [1,0,0],
    [1,1,1],
]
for i in range(0,200):
    err = 0
    dt = data[::]
    random.shuffle(dt)
    for j in dt:
        err += model.train(j[0],j[1],j[2],0.1)
    print(err)

print("-"*30)
for j in data:
    _, _, _, _, _, _, _, res = model.predict(j[0],j[1])
    print(j[0],",",j[1],"=",res)

代码结果：

...
0.1978539306282795
0.19794670251861882
0.19745074826953185
0.19529942727878868
0.19779970636626873
0.19661596298810918
               
0 , 0 = -0.24217968147818447
0 , 1 = 0.236033934015224
1 , 0 = 0.24457439328909888
1 , 1 = 0.5919949310028919