#一、准备
为了更深入的理解神经网络,笔者基本采用纯C++的手写方式实现,其中矩阵方面的运算则调用opencv,数据集则来自公开数据集a1a。
实验环境:
本文紧跟上篇文章深度学习实践(一)——logistic regression。
#二、神经网络基础
标准的神经网络结构如下图所示,其实就是上文logistic regression的增强版(即多加了几个隐层),基本思路还未变化。关于更详细的原理介绍,这里还是推荐吴恩达的深度学习系列课程。
下面以三层神经网络(即上图)并结合a1a数据集,介绍构建的一般步骤:
- 初始化参数w1、w2、w3和b1、b2、b3,因为a1a数据集的维度是有123个特征,所以上图中input_layer维度为(123,m),m为样本数量,如训练集则为1065;而我们所构建的三层神经网络中间隐层神经元个数分别为(64,16,1),所以初始化参数矩阵w1(123,64)、w2(64,16)、w3(16,1)和偏置实数b1、b2、b3。
- 将W和X相乘(矩阵相乘,X为上层的输出,一开始即为样本的输入),再加上偏置b(为实数),则得到Z。
- 将Z进行激活,在隐层选择激活函数relu(可以更好的防止梯度爆炸,且结果很好),输出层选择sigmoid限制输出,它们的图像如下:
- 将上面的正向传播完成后,定义损失函数,这里使用交叉熵代价函数。
- 反向传播,并更新参数。
正向传播基本公式:
这里上标L代表第几层,上标i表示第几个样本(对应到a1a数据集即第几行),如
A
[
0
]
A^{[0]}
A[0]表示0层的输入(即样本输入)。
(1)
Z
[
1
]
=
W
[
1
]
A
[
0
]
+
b
[
1
]
Z^{[1]} = W^{[1]}A^{[0]} +b^{[1]}\tag{1}
Z[1]=W[1]A[0]+b[1](1)
(2)
A
[
1
]
=
R
e
l
u
(
Z
[
1
]
)
A^{[1]} = Relu(Z^{[1]})\tag{2}
A[1]=Relu(Z[1])(2)
(3)
Z
[
2
]
=
W
[
2
]
A
[
1
]
+
b
[
2
]
Z^{[2]} = W^{[2]}A^{[1]} +b^{[2]}\tag{3}
Z[2]=W[2]A[1]+b[2](3)
(4)
A
[
2
]
=
R
e
l
u
(
Z
[
2
]
)
A^{[2]} = Relu(Z^{[2]})\tag{4}
A[2]=Relu(Z[2])(4)
(5)
Z
[
3
]
=
W
[
3
]
A
[
2
]
+
b
[
3
]
Z^{[3]} = W^{[3]}A^{[2]} +b^{[3]}\tag{5}
Z[3]=W[3]A[2]+b[3](5)
(6)
A
[
3
]
=
S
i
g
m
o
i
d
(
Z
[
3
]
)
A^{[3]} = Sigmoid(Z^{[3]})\tag{6}
A[3]=Sigmoid(Z[3])(6)
(7)
L
(
A
[
3
]
,
Y
^
)
=
−
A
[
3
]
log
(
A
[
3
]
)
−
(
1
−
Y
^
)
log
(
1
−
A
[
3
]
)
\mathcal{L}(A^{[3]}, \hat Y) = - A^{[3]}\log(A^{[3]}) - (1-\hat Y ) \log(1-A^{[3]})\tag{7}
L(A[3],Y^)=−A[3]log(A[3])−(1−Y^)log(1−A[3])(7)
The cost is then computed by summing over all training examples:
(8)
J
=
1
m
∑
i
=
1
m
L
(
A
(
i
)
[
3
]
,
Y
(
i
)
)
J = \frac{1}{m} \sum_{i=1}^m \mathcal{L}(A^{(i)[3]}, Y^{(i)})\tag{8}
J=m1i=1∑mL(A(i)[3],Y(i))(8)
反向传播基本公式:
(1)
d
A
[
3
]
=
∂
L
∂
A
[
3
]
=
1
−
Y
^
1
−
A
[
3
]
−
Y
^
A
[
3
]
dA^{[3]}= \frac{\partial \mathcal{L}}{\partial A^{[3]}}= \frac{1-\hat Y}{1-A^{[3]}}-\frac{\hat Y}{A^{[3]}}\tag{1}
dA[3]=∂A[3]∂L=1−A[3]1−Y^−A[3]Y^(1)
(2)
d
Z
[
3
]
=
∂
L
∂
A
[
3
]
∗
∂
A
[
3
]
∂
Z
[
3
]
=
d
A
[
3
]
∗
A
[
3
]
∗
(
1
−
A
[
3
]
)
dZ^{[3]}=\frac{\partial \mathcal{L}}{\partial A^{[3]}}*\frac{\partial A^{[3]} }{\partial Z^{[3]}}=dA[3]*A^{[3]}*(1-A^{[3]})\tag{2}
dZ[3]=∂A[3]∂L∗∂Z[3]∂A[3]=dA[3]∗A[3]∗(1−A[3])(2)
(3)
d
W
[
3
]
=
∂
L
∂
Z
[
3
]
∗
∂
Z
[
3
]
∂
W
[
3
]
=
1
m
d
Z
[
3
]
A
[
2
]
T
dW^{[3]} = \frac{\partial \mathcal{L} }{\partial Z^{[3]}} *\frac{\partial Z^{[3]} }{\partial{W^{[3]}}}= \frac{1}{m} dZ^{[3]} A^{[2] T} \tag{3}
dW[3]=∂Z[3]∂L∗∂W[3]∂Z[3]=m1dZ[3]A[2]T(3)
(4)
d
b
[
3
]
=
∂
L
∂
Z
[
3
]
∗
∂
Z
[
3
]
∂
b
[
3
]
=
1
m
∑
i
=
1
m
d
Z
[
3
]
(
i
)
db^{[3]} = \frac{\partial \mathcal{L} }{\partial Z^{[3]}} *\frac{\partial Z^{[3]} }{\partial{b^{[3]}}} = \frac{1}{m} \sum_{i = 1}^{m} dZ^{[3](i)}\tag{4}
db[3]=∂Z[3]∂L∗∂b[3]∂Z[3]=m1i=1∑mdZ[3](i)(4)
(5)
d
A
[
2
]
=
∂
L
∂
Z
[
3
]
∗
∂
Z
[
3
]
∂
A
[
2
]
=
W
[
3
]
T
d
Z
[
3
]
dA^{[2]} = \frac{\partial \mathcal{L} }{\partial Z^{[3]}} *\frac{\partial Z^{[3]}}{\partial A^{[2]}} = W^{[3] T} dZ^{[3]} \tag{5}
dA[2]=∂Z[3]∂L∗∂A[2]∂Z[3]=W[3]TdZ[3](5)
(6)
d
Z
[
2
]
=
∂
L
∂
A
[
2
]
∗
∂
A
[
2
]
∂
Z
[
2
]
=
d
A
[
2
]
∗
(
A
[
2
]
>
0
)
dZ^{[2]}=\frac{\partial \mathcal{L}}{\partial A^{[2]}}*\frac{\partial A^{[2]} }{\partial Z^{[2]}}=dA^{[2]}*(A^{[2]}>0)\tag{6}
dZ[2]=∂A[2]∂L∗∂Z[2]∂A[2]=dA[2]∗(A[2]>0)(6)
(7)
d
W
[
2
]
=
∂
L
∂
Z
[
2
]
∗
∂
Z
[
2
]
∂
W
[
2
]
=
1
m
d
Z
[
2
]
A
[
1
]
T
dW^{[2]} = \frac{\partial \mathcal{L} }{\partial Z^{[2]}} *\frac{\partial Z^{[2]} }{\partial{W^{[2]}}}= \frac{1}{m} dZ^{[2]} A^{[1] T} \tag{7}
dW[2]=∂Z[2]∂L∗∂W[2]∂Z[2]=m1dZ[2]A[1]T(7)
(8)
d
b
[
2
]
=
∂
L
∂
Z
[
2
]
∗
∂
Z
[
2
]
∂
b
[
2
]
=
1
m
∑
i
=
1
m
d
Z
[
2
]
(
i
)
db^{[2]} = \frac{\partial \mathcal{L} }{\partial Z^{[2]}} *\frac{\partial Z^{[2]} }{\partial{b^{[2]}}} = \frac{1}{m} \sum_{i = 1}^{m} dZ^{[2](i)}\tag{8}
db[2]=∂Z[2]∂L∗∂b[2]∂Z[2]=m1i=1∑mdZ[2](i)(8)
(9)
d
A
[
1
]
=
∂
L
∂
Z
[
2
]
∗
∂
Z
[
2
]
∂
A
[
1
]
=
W
[
2
]
T
d
Z
[
2
]
dA^{[1]} = \frac{\partial \mathcal{L} }{\partial Z^{[2]}} *\frac{\partial Z^{[2]}}{\partial A^{[1]}} = W^{[2] T} dZ^{[2]} \tag{9}
dA[1]=∂Z[2]∂L∗∂A[1]∂Z[2]=W[2]TdZ[2](9)
(10)
d
Z
[
1
]
=
∂
L
∂
A
[
1
]
∗
∂
A
[
1
]
∂
Z
[
1
]
=
d
A
[
1
]
∗
(
A
[
1
]
>
0
)
dZ^{[1]}=\frac{\partial \mathcal{L}}{\partial A^{[1]}}*\frac{\partial A^{[1]} }{\partial Z^{[1]}}=dA^{[1]}*(A^{[1]}>0)\tag{10}
dZ[1]=∂A[1]∂L∗∂Z[1]∂A[1]=dA[1]∗(A[1]>0)(10)
(11)
d
W
[
1
]
=
∂
L
∂
Z
[
1
]
∗
∂
Z
[
1
]
∂
W
[
1
]
=
1
m
d
Z
[
1
]
A
[
0
]
T
dW^{[1]} = \frac{\partial \mathcal{L} }{\partial Z^{[1]}} *\frac{\partial Z^{[1]} }{\partial{W^{[1]}}}= \frac{1}{m} dZ^{[1]} A^{[0] T} \tag{11}
dW[1]=∂Z[1]∂L∗∂W[1]∂Z[1]=m1dZ[1]A[0]T(11)
(12)
d
b
[
1
]
=
∂
L
∂
Z
[
1
]
∗
∂
Z
[
1
]
∂
b
[
1
]
=
1
m
∑
i
=
1
m
d
Z
[
1
]
(
i
)
db^{[1]} = \frac{\partial \mathcal{L} }{\partial Z^{[1]}} *\frac{\partial Z^{[1]} }{\partial{b^{[1]}}} = \frac{1}{m} \sum_{i = 1}^{m} dZ^{[1](i)}\tag{12}
db[1]=∂Z[1]∂L∗∂b[1]∂Z[1]=m1i=1∑mdZ[1](i)(12)
#三、实践
数据集介绍、处理及一些公用的函数已在系列的上一篇文章,故在此不做赘述(只写出函数声明)。
从文件中创建矩阵:
void creatMat(Mat &x, Mat &y, String fileName);
初始化参数(这里使用xavier初始化):
void initial_parermaters(Mat &w, double &b, int n1, int n0) {
w = Mat::zeros(n1, n0, CV_64FC1);
b = 0.0;
//double temp = 2 / (sqrt(n1));
double temp = sqrt(6.0 / (double)(n1 + n0));
RNG rng;
for (int i = 0; i < w.rows; i++) {
for (int j = 0; j < w.cols; j++) {
w.at<double>(i, j) = rng.uniform(-temp, temp);//xavier初始化
//w.at<double>(i, j) = 0;
}
}
}
relu函数的编写:
void relu(const Mat &original, Mat &response) {
response = original.clone();//防止维度不同
for (int i = 0; i < original.rows; i++) {
for (int j = 0; j < original.cols; j++) {
if (original.at<double>(i, j) < 0) {
response.at<double>(i, j) = 0.0;
}
}
}
}
正向传播:
void linear_activation_forward(Mat &a_prev, Mat &a, Mat &w, double &b, string activation) {
cv::Mat z;
if (activation == "sigmoid") {
z = (w*a_prev) + b;
//cout << w.rows<<","<<w.cols<<" " << a_prev.rows<<","<<a_prev.cols<<endl;
sigmoid(z, a);
}
else if (activation == "relu") {
z = (w*a_prev) + b;
//cout << w.rows << "," << w.cols << " " << a_prev.rows << "," << a_prev.cols << endl;
relu(z, a);
}
}
反向传播:
void activation_backward(const Mat &a, const Mat &da, Mat &dz, string activation) {
if (activation == "sigmoid") {
dz = da.mul(a.mul(1 - a));
}
else if (activation == "relu") {
dz = da.clone();//保证维度相同
for (int i = 0; i < a.rows; i++) {
for (int j = 0; j < a.cols; j++) {
if (a.at<double>(i, j) <= 0) {
dz.at<double>(i, j) = 0.0;
}
}
}
}
}
void linear_backward(const Mat &da, const Mat &a, const Mat &a_prev, Mat &w, double &b, Mat &dw, double &db, Mat &da_prev, const int m, const double learning_rate, string activation) {
cv::Mat dz;
activation_backward(a, da, dz, activation);//激活函数的反向传播
dw = (1.0 / m)*dz*a_prev.t();
db = (1.0 / m)*sum(dz)[0];
da_prev = w.t()*dz;
w = w - (learning_rate * dw);
b = b - (learning_rate * db);
}
#四、实验结果分析
迭代8000次cost分析:
我们容易发现更高的学习率可以获得较低的cost值,但是当其迭代到一定次数时,会有一定的起伏。
迭代8000次,准确率分析:
通过上图,易发现在一定迭代次数后,训练集和测试集准确率都会产生起伏,而且当训练集准确率不断上升时,测试集却未增长反而下降,最终产生了过拟合现象。
#五、结语
神经网络层数深,参数多,所以很难训练,一般训练8000次所需时间很长,下面的一篇文章主要讲一些优化方法(如adam),及如何处理过拟合(如dropout)等。
实验地址:码云