www.gusucode.com > 有监督的 CNN 网络完成对MNIST 数字的识别 > 有监督的 CNN 网络完成对MNIST 数字的识别/CNN—卷积神经网络数字识别/@cnn/cnn.m
function cnet = cnn(numLayers,numFLayers,numInputs,InputWidth,InputHeight,numOutputs)
%CNN convolutional neural network class constructor
%
% Syntax
%
% cnet =
% cnn(numLayers,numFLayers,numInputs,InputWidth,InputHeight,numOutputs)
%
% Description
% Input:
% numLayers - total number of layers
% numFLayers - number of fully connected layers
% numInputs - number of input images (currently only 1 supported)
% InputWidth - input image width
% InputHeight - input image heigth
% numOutputs - number of outputs
% Output:
% cnet - convolutional neural network class object
%
% Semantic is quite simple: subsampling and convolutional layers are
% follows in pairs and called SLayers and CLayers. Thus all S-layers are
% odd and all C-layers are even. After last CLayer follows FLayer wich is
% fully connected layer. The same way named weights and biases.
% Example of accessing weights:
% cnn.CLayer{2}.WC
% cnn.SLayer{3}.BS
% If it necessary to create network with first CLayer make the SLayer{1}
% linear
%Create empty network
%----User defined parameters
if(nargin<6) %If no parameters are defined set it to defaults
if((nargin==1)&&(isstruct(numLayers)))
cnet = class(numLayers,'cnn');
return;
else
cnet.numLayers = 3; %Total layers number
cnet.numSLayers = 1; %Number of S-layers
cnet.numCLayers = 1; %Number of C-layers
cnet.numFLayers = 1; %Number of F-lauers
cnet.numInputs = 1; %Number of input images
cnet.InputWidth = 30; %Input weight
cnet.InputHeight = 30; %Inout height
cnet.numOutputs = 1; %Outputs number
end
else
cnet.numLayers = numLayers;
cnet.numSLayers = ceil((numLayers-numFLayers)/2);
cnet.numCLayers = numLayers-numFLayers-cnet.numSLayers;
cnet.numFLayers = numFLayers;
cnet.numInputs = numInputs;
cnet.InputWidth = InputWidth;
cnet.InputHeight = InputHeight;
cnet.numOutputs = numOutputs;
end
%Default parameters wich typically redefined later
cnet.Perf = 'mse'; %Performance function
cnet.mu = 0.01; %Mu coefficient for stochastic Levenberg-Marqwardt
cnet.mu_dec = 0.1; %Mu per epoch decrease rate
cnet.mu_inc = 10; %Mu per epoch increase rate
cnet.mu_max = 1.0000e+010; %Maximum mu
cnet.epochs = 50; %Number of epochs
cnet.goal = 0.00001; %Goal RMSE value
cnet.teta = 0.2;
cnet.teta_dec = 0.3; %Teta per epoch decrease rate
%The way Hessian diagonal approximation is computed
%0 - Hessian running estimate is calculated every iteration
%1 - Hessian approximation is recalculated every cnet.Hrecomp iterations
%2 - No Hessian calculations are made. Pure stochastic gradient
cnet.HcalcMode = 0;
cnet.Hrecalc = 1000; %Number of iterations to passs for Hessian recalculation
cnet.HrecalcSamplesNum = 100; %Number of samples for Hessian recalculation
%Train plot properties
cnet.MCRrecalc = 200; %How often to recalculate missclassification rate
cnet.MCRsamples = 70; %How much test samples to use for MCR calculation
cnet.RMSErecalc = 10; %How often to recalculate RMSE
%SLayer contains information about subsampling layers
%Constructor only set default values for all variables
%All these variables has to be set and then Init method called
for k=1:cnet.numSLayers
m=2*k-1; %Use m to consider layer parity
%----User defined parameters
cnet.SLayer{m}.teta = 0.2; %Layer train coefficient
cnet.SLayer{m}.SRate = 2; %Subsampling rate
cnet.SLayer{m}.TransfFunc = 'tansig_mod'; %Activation function
%----Feature maps, calculating while simulation of network
cnet.SLayer{m}.YS{1} = 0; %Weighted inputs (before activation function)
cnet.SLayer{m}.XS{1} = 0; %Outputs (after activation)
cnet.SLayer{m}.SS{1} = 0; %Subsampled feature map
%----Parameters initialized by Init method
cnet.SLayer{m}.WS{1} = 0; %Weights
cnet.SLayer{m}.BS{1} = 0; %Biases
cnet.SLayer{m}.numFMaps = 1; %Number of output feature maps
cnet.SLayer{m}.FMapWidth = 10; %Feature maps dimensions
cnet.SLayer{m}.FMapHeight = 10;
cnet.SLayer{m}.ln = m; %Layer number
%----Variables calculated while training
cnet.SLayer{m}.dEdW{1} = 0; %Partial derivative of error by weights
cnet.SLayer{m}.dEdB{1} = 0; %Partial derivative of error by biases
cnet.SLayer{m}.dEdX{1} = 0; %Partial derivative of error by outputs
cnet.SLayer{m}.dXdY{1} = 0; %Partial derivative of outputs by weighted inputs
cnet.SLayer{m}.dYdW{1} = 0; %Partial derivative of weighted inputs by weights
cnet.SLayer{m}.dYdB{1} = 0; %Partial derivative of weighted inputs by biases
cnet.SLayer{m}.H{1} = 0; %Hessian approximation
cnet.SLayer{m}.mu = 0; %Regularisation factor.
cnet.SLayer{m}.dEdX_last{1} = 0;
end
%CLayer - convolutional layer
for k=1:cnet.numCLayers
m=k*2;
%----User defined parameters
cnet.CLayer{m}.teta = 0.2; %暑?翳鲨屙?钺篦屙? 潆?耠?
cnet.CLayer{m}.numKernels = 1; %暑腓麇耱忸 ?屦 疋屦蜿?(蝾 驽, 黩?lenght(WC))
cnet.CLayer{m}.KernWidth = 3; %朽珈屦眍耱??疣 疋屦蜿?
cnet.CLayer{m}.KernHeight = 3 ;
%----Feature maps, calculating while simulation of network
cnet.CLayer{m}.YC = cell(1);
cnet.CLayer{m}.XC = cell(1);
%----Parameters initialized by Init method
cnet.CLayer{m}.WC{1} = 0;
cnet.CLayer{m}.BC{1} = 0;
cnet.CLayer{m}.numFMaps = 1;
cnet.CLayer{m}.FMapWidth = 10;
cnet.CLayer{m}.FMapHeight = 10;
cnet.CLayer{m}.ln = m;
%----Variables calculated while training
cnet.CLayer{m}.dEdW{1} = 0;
cnet.CLayer{m}.dEdB{1} = 0;
cnet.CLayer{m}.dEdX{1} = 0;
cnet.CLayer{m}.dXdY{1} = 0;
cnet.CLayer{m}.dYdW{1} = 0;
cnet.CLayer{m}.dYdB{1} = 0;
cnet.CLayer{m}.H{1} = 0;
cnet.CLayer{m}.mu = 0;
cnet.CLayer{m}.dEdX_last{1} = 0;
%Connection map - row number corresponds to output, column number
%corresponds to input. Necessary for network assymetry
cnet.CLayer{m}.ConMap = 0;
end
%FLayer - fully-connected layer
for k=cnet.numCLayers+cnet.numSLayers+1:cnet.numLayers
%----User defined parameters
cnet.FLayer{k}.teta = 0.2;
if k==cnet.numLayers
cnet.FLayer{k}.numNeurons = cnet.numOutputs; %If the layer is output
else
cnet.FLayer{k}.numNeurons = 10; %Default number of neurons in layer
end
cnet.FLayer{k}.W = 0;
cnet.FLayer{k}.B = 0;
%----Feature maps, calculating while simulation of network
cnet.FLayer{k}.Y = 0;
cnet.FLayer{k}.X = 0;
cnet.FLayer{k}.ln = k;
cnet.FLayer{k}.TransfFunc = 'tansig_mod';
%----Variables calculated while training
cnet.FLayer{k}.dEdW{1} = 0;
cnet.FLayer{k}.dEdB{1} = 0;
cnet.FLayer{k}.dEdX{1} = 0;
cnet.FLayer{k}.dXdY{1} = 0;
cnet.FLayer{k}.dYdW{1} = 0;
cnet.FLayer{k}.dYdB{1} = 0;
cnet.FLayer{k}.H{1} = 0;
cnet.FLayer{k}.mu = 0;
cnet.FLayer{k}.dEdX_last{1} = 0;
end
cnet = class(cnet,'cnn');