BP算法的实现步骤

BP算法实现步骤（软件）：

1）初始化

2）输入训练样本对，计算各层输出

3）计算网络输出误差

4）计算各层误差信号

5）调整各层权值

6）检查网络总误差是否达到精度要求

满足，则训练结束；不满足，则返回步骤2）

3、多层感知器（基于BP算法）的主要能力：

1）非线性映射：足够多样本－>学习训练

能学习和存储大量输入－输出模式映射关系。只要能提供足够多的样本模式对供BP网络进行学习训练，它便能完成由n维输入空间到m维输出空间的非线橡液性映射。

2）泛化：输入梁肆物新样本（训练时未有）－>完成正确的输入、输出映射

3）容错：个别样本误差不能左右对权矩阵的调整

4、标准BP算法的缺陷：

1）易形成局部极小（属贪婪算法，局部最优)而得不到全局最优；

2）训练次数多使得学习效率低下，收敛速度慢（需做大量运算）；

3）隐节点的选取缺乏理论支持；

4）训练时学习新样本有遗忘旧样本趋势。

注3：改进算法—增加动量项、自适应调整学习速率（这个似乎不错)及引入雹春陡度因子

function Solar_SAE

tic

n = 300

m=20

train_x = []

test_x = []

for i = 1:n

%filename = strcat(['D:\Program Files\MATLAB\型此R2012a\work\DeepLearn\Solar_SAE\64_64_3train\' num2str(i,'%03d') '.bmp'])

%filename = strcat(['E:\matlab\work\c0\TrainImage' num2str(i,'%03d') '.bmp'])

filename = strcat(['E:\image restoration\3-(' num2str(i) ')-4.jpg'])

b = imread(filename)

%c = rgb2gray(b)

c=b

[ImageRow ImageCol] = size(c)

c = reshape(c,[1,ImageRow*ImageCol])

train_x = [train_xc]

end

for i = 1:m

%filename = strcat(['D:\Program Files\MATLAB\R2012a\work\DeepLearn\Solar_SAE\64_64_3test\' num2str(i,'%03d') '嫌租盯.bmp'])

%filename = strcat(['E:\matlab\work\c0\TestImage'芹和 num2str(i+100,'%03d') '-1.bmp'])

filename = strcat(['E:\image restoration\3-(' num2str(i+100) ').jpg'])

b = imread(filename)

%c = rgb2gray(b)

c=b

[ImageRow ImageCol] = size(c)

c = reshape(c,[1,ImageRow*ImageCol])

test_x = [test_xc]

end

train_x = double(train_x)/255

test_x = double(test_x)/255

%train_y = double(train_y)

%test_y = double(test_y)

% Setup and train a stacked denoising autoencoder (SDAE)

rng(0)

%sae = saesetup([4096 500 200 50])

%sae.ae{1}.activation_function = 'sigm'

%sae.ae{1}.learningRate = 0.5

%sae.ae{1}.inputZeroMaskedFraction = 0.0

%sae.ae{2}.activation_function = 'sigm'

%sae.ae{2}.learningRate = 0.5

%%sae.ae{2}.inputZeroMaskedFraction = 0.0

%sae.ae{3}.activation_function = 'sigm'

%sae.ae{3}.learningRate = 0.5

%sae.ae{3}.inputZeroMaskedFraction = 0.0

%sae.ae{4}.activation_function = 'sigm'

%sae.ae{4}.learningRate = 0.5

%sae.ae{4}.inputZeroMaskedFraction = 0.0

%opts.numepochs = 10

%opts.batchsize = 50

%sae = saetrain(sae, train_x, opts)

%visualize(sae.ae{1}.W{1}(:,2:end)')

% Use the SDAE to initialize a FFNN

nn = nnsetup([4096 1500 500 200 50 200 500 1500 4096])

nn.activation_function = 'sigm'

nn.learningRate = 0.03

nn.output = 'linear'% output unit 'sigm' (=logistic), 'softmax' and 'linear'

%add pretrained weights

%nn.W{1} = sae.ae{1}.W{1}

%nn.W{2} = sae.ae{2}.W{1}

%nn.W{3} = sae.ae{3}.W{1}

%nn.W{4} = sae.ae{3}.W{2}

%nn.W{5} = sae.ae{2}.W{2}

%nn.W{6} = sae.ae{1}.W{2}

%nn.W{7} = sae.ae{2}.W{2}

%nn.W{8} = sae.ae{1}.W{2}

% Train the FFNN

opts.numepochs = 30

opts.batchsize = 150

tx = test_x(14,:)

nn1 = nnff(nn,tx,tx)

ty1 = reshape(nn1.a{9},64,64)

nn = nntrain(nn, train_x, train_x, opts)

toc

tic

nn2 = nnff(nn,tx,tx)

toc

tic

ty2 = reshape(nn2.a{9},64,64)

tx = reshape(tx,64,64)

tz = tx - ty2

tz = im2bw(tz,0.1)

%imshow(tx)

%figure,imshow(ty2)

%figure,imshow(tz)

ty = cat(2,tx,ty2,tz)

montage(ty)

filename3 = strcat(['E:\image restoration\3.jpg'])

e=imread(filename3)

f= rgb2gray(e)

f=imresize(f,[64,64])

%imshow(ty2)

f=double (f)/255

[PSNR, MSE] = psnr(ty2,f)

imwrite(ty2,'E:\image restoration\bptest.jpg','jpg')

toc

%visualize(ty)

%[er, bad] = nntest(nn, tx, tx)

%assert(er <0.1, 'Too big error')

%读取训练数据

[f1,f2,f3,f4,class] = textread('trainData.txt' , '%f%f%f%f%f',150)

%特征值归一化

[input,minI,maxI] = premnmx( [f1 , f2 , f3 , f4 ]')

%构造输出矩阵

s = length( class)

output = zeros( s , 3 )

for i = 1 : s

output( i , class( i ) ) = 1

end

%创建神慎圆绝宽姿经网络

net = newff( minmax(input) , [10 3] , { 'logsig' 'purelin' } , 'traingdx' )

%设置训练参数

net.trainparam.show = 50

net.trainparam.epochs = 500

net.trainparam.goal = 0.01

net.trainParam.lr = 0.01

%开始训练

net = train( net, input , output' )

%读取测试数据

[t1 t2 t3 t4 c] = textread('testData.txt' , '%f%f%f%f%f',150)

%测试数据归腔誉一化

testInput = tramnmx ( [t1,t2,t3,t4]' , minI, maxI )

%仿真

Y = sim( net , testInput )

%统计识别正确率

[s1 , s2] = size( Y )

hitNum = 0

for i = 1 : s2

[m , Index] = max( Y( : , i ) )

if( Index == c(i) )

hitNum = hitNum + 1

end

end

sprintf('识别率是 %3.3f%%',100 * hitNum / s2 )

看了你的数据，你至少要有的类标号吧，不知道你哪里是输入向量，哪里是输出向量

---