学步园

性能分析代码如下：

% charRecogonize.m<br /> clc;<br /> clear;<br /> %调用prodat产生alpabet字母表和targets，target为eye(26);<br /> [alphabet,targets] = prodat;<br /> P =alphabet; % alphabet = [ letterA, letterB, ..., letterZ ];<br /> T = targets; % targets = eye(26);<br /> [R,Q] = size(P);<br /> % 神经元个数<br /> S1 = 10;<br /> % 求T的大小，此时他的列数与P的列数为一致<br /> [S2,Q] = size(T);<br /> %创建网络<br /> net = newff(minmax(alphabet),[S1 S2],{'logsig' 'logsig'},'traingdx');<br /> net.LW{2,1} = net.LW{2,1}*0.01;<br /> net.b{2} = net.b{2}*0.01;<br /> % 无噪声字母识别网络的训练<br /> fprintf('================================================/n');<br /> fprintf('There are first train,begin/n');<br /> net.performFcn = 'sse'; %执行函数为误差平方和函数<br /> net.trainParam.epochs=800; %最大训练步长<br /> net.trainParam.lr=0.01; %学习率<br /> net.trainParam.lr_inc =1.05; %增长的学习率<br /> net.trainParam.lr_dec =0.7;<br /> net.trainParam.goal=0.1; %执行函数目标值<br /> net.trainParam.mc =0.9; %附加动量因子<br /> net.trainParam.min_grad=1e-10; %最小执行梯度<br /> net.trainParam.show = 50;<br /> % P = alphabet;<br /> % T = targets;<br /> [net,tr] = train(net,P,T);<br /> fprintf('There are first train,end/n');<br /> fprintf('================================================/n');<br /> % noisyJ = alphabet(:,25)+randn(35,1) * 0.2;<br /> % A2 = sim(net,noisyJ);<br /> % A2 = compet(A2);<br /> % compet is a transfer function. Transfer functions calculate a layer's output from its net input.<br /> % compet(N) takes one input argument,<br /> % N - S x Q matrix of net input (column) vectors.<br /> % and returns output vectors with 1 where each net input vector has its maximum value, and 0 elsewhere<br /> % answer = find(compet(A2) == 1);<br /> % indices = find(X) returns the linear indices corresponding to the nonzero entries of the array X.<br /> % If none are found, find returns an empty, 0-by-1 matrix.<br /> % In general, find(X) regards X as X(:),<br /> % which is the long column vector formed by concatenating the columns of X.<br /> % Some operations on a vector x = [11 0 33 0 55]';<br /> % find(x)<br /> %<br /> % ans =<br /> %<br /> % 1<br /> % 3<br /> % 5<br /> %<br /> % find(x == 0)<br /> %<br /> % ans =<br /> %<br /> % 2<br /> % 4<br /> %<br /> % find(0 < x & x < 10*pi)<br /> %<br /> % ans =<br /> %<br /> % 1<br /> % plotchar(alphabet(:,answer));<br /> % figure(2);<br /> % plotchar(noisyJ);<br /> %%%%%%%%%%%%%%%=========================================================<br /> %%%%%%%%%%% 具有噪声的输入识别网络的训练，共循环10次，开始<br /> fprintf('================================================/n');<br /> fprintf('There are 10 times,For noise,begin/n');<br /> netn = net;<br /> netn.trainParam.goal = 0.1; %目标误差.<br /> netn.trainParam.epochs = 300; %最大训练步长.<br /> for pass=1:10<br /> fprintf('Pass = %.0f/n',pass);<br /> T = [targets targets targets targets];<br /> P = [alphabet, alphabet, ...<br /> (alphabet + randn(R,Q)*0.1), ...<br /> (alphabet + randn(R,Q)*0.2)];<br /> [netn,tr] = train(netn,P,T);<br /> end<br /> fprintf('There are 10 times,For noise,end/n');<br /> fprintf('================================================/n');<br /> %%%%%%%%%%% 具有噪声的输入识别网络的训练，共循环10次，结束</p> <p>%%%%%%% 再次用无噪声字母训练识别网络，开始<br /> fprintf('================================================/n');<br /> fprintf('There are no noise,begin/n');<br /> netn.trainParam.goal = 0.1; % 均方误差目标.<br /> netn.trainParam.epochs = 500; % 最大训练步长<br /> net.trainParam.show = 5; % 训练中显示的频率<br /> P = alphabet;<br /> T = targets;<br /> [netn,tr] = train(netn,P,T);<br /> %%%%%%% 再次用无噪声字母训练识别网络，结束<br /> fprintf('There are no noise,end/n');<br /> fprintf('================================================/n');<br /> %%%% 测试所训练的网络性能<br /> noise_range = 0:.02:0.2;<br /> max_test = 100;<br /> network1 = [];<br /> network2 = [];<br /> T = targets;<br /> % PERFORM THE TEST<br /> for noiselevel = noise_range<br /> fprintf('================================================/n');<br /> fprintf('Testing networks with noise level of %.2f./n',noiselevel);</p> <p> errors1 = 0;<br /> errors2 = 0;<br /> for i=1:max_test<br /> P = alphabet + randn(35,26)*noiselevel;<br /> % TEST NETWORK 1<br /> A = sim(net,P);<br /> AA = compet(A);<br /> errors1 = errors1 + sum(sum(abs(AA-T)))/2;<br /> % TEST NETWORK 2<br /> An = sim(netn,P);<br /> AAn = compet(An);<br /> errors2 = errors2 + sum(sum(abs(AAn-T)))/2;<br /> echo off<br /> end<br /> % 100组26个矢量的平均误差<br /> network1 = [network1 errors1/26/100];<br /> network2 = [network2 errors2/26/100];<br /> end<br /> fprintf('================================================/n');<br /> % echo on<br /> %<br /> % The echo command controls the echoing of M-files during execution.<br /> % Normally, the commands in M-files are not displayed on the screen<br /> % during execution. Command echoing is useful for debugging<br /> % or for demonstrations, allowing the commands to be viewed as they execute.<br /> % The echo command behaves in a slightly different manner for script files<br /> % and function files. For script files, the use of echo is simple;<br /> % echoing can be either on or off, in which case any script used is affected.<br /> % echo onTurns on the echoing of commands in all script filesecho offTurns off the echoing of commands in all script filesechoToggles the echo state<br /> % pause % 按任意键显示结果<br /> % 显示结果<br /> % ========<br /> clf<br /> % clf deletes from the current figure all graphics objects<br /> % whose handles are not hidden (i.e., their HandleVisibility property is set to on).<br /> % clf('reset') deletes from the current figure all graphics objects regardless of<br /> % the setting of their HandleVisibility property and resets all figure properties except Position,<br /> % Units, PaperPosition, and PaperUnits to their default values.<br /> % figure_handle = clf(...) return the handle of the figure.<br /> % This is useful when the figure IntegerHandle property is off<br /> % since the noninteger handle becomes invalid when the reset option is used<br /> % (i.e., IntegerHandle is reset to on, which is the default).<br /> plot(noise_range,network1*100,'r*',noise_range,network2*100,'g*');<br /> grid<br /> xlabel('Noise Level');<br /> ylabel('Percentage of Recognition');<br /> % echo off<br /> %三种情况下字母Y的表示情况<br /> noisyY1 = alphabet(:,25)+randn(35,1) * 0.2;<br /> figure(2);<br /> plotchar(noisyY1);<br /> noiseY2=alphabet(:,25)+randn(35,1)*0.1;<br /> figure(3);<br /> plotchar(noiseY2);<br /> idealY=alphabet(:,25);<br /> figure(4)<br /> plotchar(idealY);<br /> %%%%%%%%%%%%%%%=========================================================<br /> % figure(5);<br /> % sp=sim(net,noisyY1);<br /> % plotchar(sp);<br /> % figure(6);<br /> % sp2=sim(net,noisyY1);<br /> % plotchar(sp2);<br />

调用的prodat，预处理数据用于产生字母序列表和eye矩阵，具体程序实现如下：

% prodat.m recognize.m调用的字母产生程序<br /> function[alphabet,targets] = prodat()<br /> % Returns:<br /> % ALPHABET - 35x26 matrix of 5x7 bit maps for each letter.<br /> % TARGETS - 26x26 target vectors. </p> <p>letterA = [0 0 1 0 0 ...<br /> 0 1 0 1 0 ...<br /> 0 1 0 1 0 ...<br /> 1 0 0 0 1 ...<br /> 1 1 1 1 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ]'; </p> <p>letterB = [1 1 1 1 0 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 1 1 1 0 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 1 1 1 0 ]'; </p> <p>letterC = [0 1 1 1 0 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 0 ...<br /> 1 0 0 0 0 ...<br /> 1 0 0 0 0 ...<br /> 1 0 0 0 1 ...<br /> 0 1 1 1 0 ]'; </p> <p>letterD = [1 1 1 1 0 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 1 1 1 0 ]'; </p> <p>letterE = [1 1 1 1 1 ...<br /> 1 0 0 0 0 ...<br /> 1 0 0 0 0 ...<br /> 1 1 1 1 0 ...<br /> 1 0 0 0 0 ...<br /> 1 0 0 0 0 ...<br /> 1 1 1 1 1 ]'; </p> <p>letterF = [1 1 1 1 1 ...<br /> 1 0 0 0 0 ...<br /> 1 0 0 0 0 ...<br /> 1 1 1 1 0 ...<br /> 1 0 0 0 0 ...<br /> 1 0 0 0 0 ...<br /> 1 0 0 0 0 ]'; </p> <p>letterG = [0 1 1 1 0 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 0 ...<br /> 1 0 0 0 0 ...<br /> 1 0 0 1 1 ...<br /> 1 0 0 0 1 ...<br /> 0 1 1 1 0 ]'; </p> <p>letterH = [1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 1 1 1 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ]'; </p> <p>letterI = [0 1 1 1 0 ...<br /> 0 0 1 0 0 ...<br /> 0 0 1 0 0 ...<br /> 0 0 1 0 0 ...<br /> 0 0 1 0 0 ...<br /> 0 0 1 0 0 ...<br /> 0 1 1 1 0 ]'; </p> <p>letterJ = [1 1 1 1 1 ...<br /> 0 0 1 0 0 ...<br /> 0 0 1 0 0 ...<br /> 0 0 1 0 0 ...<br /> 0 0 1 0 0 ...<br /> 1 0 1 0 0 ...<br /> 0 1 0 0 0 ]'; </p> <p>letterK = [1 0 0 0 1 ...<br /> 1 0 0 1 0 ...<br /> 1 0 1 0 0 ...<br /> 1 1 0 0 0 ...<br /> 1 0 1 0 0 ...<br /> 1 0 0 1 0 ...<br /> 1 0 0 0 1 ]'; </p> <p>letterL = [1 0 0 0 0 ...<br /> 1 0 0 0 0 ...<br /> 1 0 0 0 0 ...<br /> 1 0 0 0 0 ...<br /> 1 0 0 0 0 ...<br /> 1 0 0 0 0 ...<br /> 1 1 1 1 1 ]'; </p> <p>letterM = [1 0 0 0 1 ...<br /> 1 1 0 1 1 ...<br /> 1 0 1 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ]'; </p> <p>letterN = [1 0 0 0 1 ...<br /> 1 1 0 0 1 ...<br /> 1 1 0 0 1 ...<br /> 1 0 1 0 1 ...<br /> 1 0 0 1 1 ...<br /> 1 0 0 1 1 ...<br /> 1 0 0 0 1 ]'; </p> <p>letterO = [0 1 1 1 0 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 0 1 1 1 0 ]'; </p> <p>letterP = [1 1 1 1 0 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 1 1 1 0 ...<br /> 1 0 0 0 0 ...<br /> 1 0 0 0 0 ...<br /> 1 0 0 0 0 ]'; </p> <p>letterQ = [0 1 1 1 0 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 1 0 1 ...<br /> 1 0 0 1 0 ...<br /> 0 1 1 0 1 ]'; </p> <p>letterR = [1 1 1 1 0 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 1 1 1 0 ...<br /> 1 0 1 0 0 ...<br /> 1 0 0 1 0 ...<br /> 1 0 0 0 1 ]'; </p> <p>letterS = [0 1 1 1 0 ...<br /> 1 0 0 0 1 ...<br /> 0 1 0 0 0 ...<br /> 0 0 1 0 0 ...<br /> 0 0 0 1 0 ...<br /> 1 0 0 0 1 ...<br /> 0 1 1 1 0 ]'; </p> <p>letterT = [1 1 1 1 1 ...<br /> 0 0 1 0 0 ...<br /> 0 0 1 0 0 ...<br /> 0 0 1 0 0 ...<br /> 0 0 1 0 0 ...<br /> 0 0 1 0 0 ...<br /> 0 0 1 0 0 ]'; </p> <p>letterU = [1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 0 1 1 1 0 ]'; </p> <p>letterV = [1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 0 1 0 1 0 ...<br /> 0 0 1 0 0 ]'; </p> <p>letterW = [1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 1 0 1 0 1 ...<br /> 1 1 0 1 1 ...<br /> 1 0 0 0 1 ]'; </p> <p>letterX = [1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 0 1 0 1 0 ...<br /> 0 0 1 0 0 ...<br /> 0 1 0 1 0 ...<br /> 1 0 0 0 1 ...<br /> 1 0 0 0 1 ]'; </p> <p>letterY = [1 0 0 0 1 ...<br /> 1 0 0 0 1 ...<br /> 0 1 0 1 0 ...<br /> 0 0 1 0 0 ...<br /> 0 0 1 0 0 ...<br /> 0 0 1 0 0 ...<br /> 0 0 1 0 0 ]'; </p> <p>letterZ = [1 1 1 1 1 ...<br /> 0 0 0 0 1 ...<br /> 0 0 0 1 0 ...<br /> 0 0 1 0 0 ...<br /> 0 1 0 0 0 ...<br /> 1 0 0 0 0 ...<br /> 1 1 1 1 1 ]'; </p> <p>alphabet = [letterA,letterB,letterC,letterD,letterE,letterF,letterG,letterH,...<br /> letterI,letterJ,letterK,letterL,letterM,letterN,letterO,letterP,...<br /> letterQ,letterR,letterS,letterT,letterU,letterV,letterW,letterX,...<br /> letterY,letterZ]; </p> <p>targets = eye(26);<br /> % Y = eye(n) returns the n-by-n identity matrix.<br /> % Y = eye(m,n) or eye([m n]) returns an m-by-n matrix with 1's on the diagonal and 0's elsewhere.<br /> % Y = eye(size(A)) returns an identity matrix the same size as A.<br /> % eye(m, n, classname) or eye([m,n],classname) is an m-by-n matrix<br /> % with 1's of class classname on the diagonal and zeros of class classname elsewhere.<br /> % classname is a string specifying the data type of the output.<br /> % classname can have the following values:<br /> % 'double', 'single', 'int8', 'uint8', 'int16', 'uint16', 'int32', or 'uint32'.

程序输出如下：

================================================<br /> There are first train,begin<br /> TRAINGDX, Epoch 0/800, SSE 170.085/0.1, Gradient 55.6444/1e-010<br /> TRAINGDX, Epoch 50/800, SSE 24.781/0.1, Gradient 0.497832/1e-010<br /> TRAINGDX, Epoch 100/800, SSE 17.3254/0.1, Gradient 0.796657/1e-010<br /> TRAINGDX, Epoch 150/800, SSE 0.153578/0.1, Gradient 0.0354647/1e-010<br /> TRAINGDX, Epoch 155/800, SSE 0.0997842/0.1, Gradient 0.0224737/1e-010<br /> TRAINGDX, Performance goal met.<br /> There are first train,end<br /> ================================================<br /> ================================================<br /> There are 10 times,For noise,begin<br /> Pass = 1<br /> TRAINGDX, Epoch 0/300, SSE 4.0244/0.1, Gradient 2.94032/1e-010<br /> TRAINGDX, Epoch 50/300, SSE 1.26719/0.1, Gradient 0.491154/1e-010<br /> TRAINGDX, Epoch 100/300, SSE 0.549706/0.1, Gradient 0.10225/1e-010<br /> TRAINGDX, Epoch 150/300, SSE 0.12426/0.1, Gradient 0.0197092/1e-010<br /> TRAINGDX, Epoch 155/300, SSE 0.0999154/0.1, Gradient 0.0158492/1e-010<br /> TRAINGDX, Performance goal met.<br /> Pass = 2<br /> TRAINGDX, Epoch 0/300, SSE 3.99564/0.1, Gradient 4.43371/1e-010<br /> TRAINGDX, Epoch 50/300, SSE 0.531749/0.1, Gradient 0.3353/1e-010<br /> TRAINGDX, Epoch 100/300, SSE 0.217027/0.1, Gradient 0.0652191/1e-010<br /> TRAINGDX, Epoch 135/300, SSE 0.0982281/0.1, Gradient 0.0200328/1e-010<br /> TRAINGDX, Performance goal met.<br /> Pass = 3<br /> TRAINGDX, Epoch 0/300, SSE 1.32628/0.1, Gradient 3.27334/1e-010<br /> TRAINGDX, Epoch 50/300, SSE 0.302672/0.1, Gradient 0.255984/1e-010<br /> TRAINGDX, Epoch 100/300, SSE 0.126619/0.1, Gradient 0.0412811/1e-010<br /> TRAINGDX, Epoch 116/300, SSE 0.0998277/0.1, Gradient 0.0253413/1e-010<br /> TRAINGDX, Performance goal met.<br /> Pass = 4<br /> TRAINGDX, Epoch 0/300, SSE 0.79332/0.1, Gradient 2.1395/1e-010<br /> TRAINGDX, Epoch 50/300, SSE 0.255685/0.1, Gradient 0.180019/1e-010<br /> TRAINGDX, Epoch 100/300, SSE 0.132418/0.1, Gradient 0.0417362/1e-010<br /> TRAINGDX, Epoch 119/300, SSE 0.099968/0.1, Gradient 0.0240296/1e-010<br /> TRAINGDX, Performance goal met.<br /> Pass = 5<br /> TRAINGDX, Epoch 0/300, SSE 1.55642/0.1, Gradient 3.23532/1e-010<br /> TRAINGDX, Epoch 50/300, SSE 0.212807/0.1, Gradient 0.193534/1e-010<br /> TRAINGDX, Epoch 100/300, SSE 0.103347/0.1, Gradient 0.0388252/1e-010<br /> TRAINGDX, Epoch 103/300, SSE 0.0985598/0.1, Gradient 0.0353377/1e-010<br /> TRAINGDX, Performance goal met.<br /> Pass = 6<br /> TRAINGDX, Epoch 0/300, SSE 2.64645/0.1, Gradient 5.82475/1e-010<br /> TRAINGDX, Epoch 50/300, SSE 0.201319/0.1, Gradient 0.207158/1e-010<br /> TRAINGDX, Epoch 95/300, SSE 0.0995744/0.1, Gradient 0.037957/1e-010<br /> TRAINGDX, Performance goal met.<br /> Pass = 7<br /> TRAINGDX, Epoch 0/300, SSE 0.873476/0.1, Gradient 2.42803/1e-010<br /> TRAINGDX, Epoch 50/300, SSE 0.260369/0.1, Gradient 0.205566/1e-010<br /> TRAINGDX, Epoch 100/300, SSE 0.116383/0.1, Gradient 0.0404004/1e-010<br /> TRAINGDX, Epoch 110/300, SSE 0.09979/0.1, Gradient 0.0296191/1e-010<br /> TRAINGDX, Performance goal met.<br /> Pass = 8<br /> TRAINGDX, Epoch 0/300, SSE 0.964115/0.1, Gradient 1.98958/1e-010<br /> TRAINGDX, Epoch 50/300, SSE 0.228885/0.1, Gradient 0.208873/1e-010<br /> TRAINGDX, Epoch 100/300, SSE 0.0987516/0.1, Gradient 0.0377656/1e-010<br /> TRAINGDX, Performance goal met.<br /> Pass = 9<br /> TRAINGDX, Epoch 0/300, SSE 0.431969/0.1, Gradient 1.24053/1e-010<br /> TRAINGDX, Epoch 50/300, SSE 0.150932/0.1, Gradient 0.129991/1e-010<br /> TRAINGDX, Epoch 93/300, SSE 0.0990331/0.1, Gradient 0.0385203/1e-010<br /> TRAINGDX, Performance goal met.<br /> Pass = 10<br /> TRAINGDX, Epoch 0/300, SSE 1.58001/0.1, Gradient 4.84543/1e-010<br /> TRAINGDX, Epoch 50/300, SSE 0.235781/0.1, Gradient 0.223338/1e-010<br /> TRAINGDX, Epoch 100/300, SSE 0.103898/0.1, Gradient 0.0375891/1e-010<br /> TRAINGDX, Epoch 103/300, SSE 0.0993984/0.1, Gradient 0.0344036/1e-010<br /> TRAINGDX, Performance goal met.<br /> There are 10 times,For noise,end<br /> ================================================<br /> ================================================<br /> There are no noise,begin<br /> TRAINGDX, Epoch 0/500, SSE 0.013799/0.1, Gradient 0.013598/1e-010<br /> TRAINGDX, Performance goal met.<br /> There are no noise,end<br /> ================================================<br /> ================================================<br /> Testing networks with noise level of 0.00.<br /> ================================================<br /> Testing networks with noise level of 0.02.<br /> ================================================<br /> Testing networks with noise level of 0.04.<br /> ================================================<br /> Testing networks with noise level of 0.06.<br /> ================================================<br /> Testing networks with noise level of 0.08.<br /> ================================================<br /> Testing networks with noise level of 0.10.<br /> ================================================<br /> Testing networks with noise level of 0.12.<br /> ================================================<br /> Testing networks with noise level of 0.14.<br /> ================================================<br /> Testing networks with noise level of 0.16.<br /> ================================================<br /> Testing networks with noise level of 0.18.<br /> ================================================<br /> Testing networks with noise level of 0.20.<br /> ================================================

运行时，多次运行显示进行模糊处理后效果更好，但是也出现了模糊处理后效果不好的情况。