基于MINST数据库的手写体数字识别CNN设计，其中CNN自己编程实现，不使用matlab工具箱-源码

function [W1,W3,W4] = DeltaMiniBatch(W1,W3,W4,X,D,N) alpha = 0.01; V1 = zeros(20,20,20); M = 5; for i = 1:(N/M) dW3sum = zeros(100,2000); dW4sum = zeros(10,100); for j = 1:M [i,j,(N/M)] id = (i-1)*M+j; x = X(:,:,id); d = D(:,id); %滤波 for epoch = 1:20 V1(:,:,epoch) = filter2(W1(:,:,epoch), x, 'valid'); end %ReLU Y1 = max(0,V1); %2×2平均池化 Y2 = (Y1(1:2:end,1:2:end,:)+Y1(2:2:end,1:2:end,:)+Y1(1:2:end,2:2:end,:)+Y1(2:2:end,2:2:end,:))/4; %Reshape变为列向量 y2 = reshape(Y2,[],1); v3 = W3*y2; %ReLU y3 = max(0,v3); v = W4*y3; %Softmax y = Softmax(v); %向前传播误差，并更新权重 e = d - y; delta = e; e3 = W4'*delta; delta3 = (v3 > 0).*e3; e2 = W3'*delta3; dW4 = alpha*delta*y3'; dW3 = alpha*delta3*y2'; E2 = reshape(e2,size(Y2)); E2_4 = E2/4; E1 = zeros(20,20,20); E1(1:2:end,1:2:end,:) = E2_4; E1(1:2:end,2:2:end,:) = E2_4; E1(2:2:end,1:2:end,:) = E2_4; E1(2:2:end,2:2:end,:) = E2_4; delta1 = (V1>0).*E1; dW1 = zeros(9,9,20); for epoch = 1:20 dW1(:,:,epoch) = alpha*filter2(delta1(:,:,epoch),x,'valid'); end W1 = W1 + dW1; dW4sum = dW4sum + dW4; dW3sum = dW3sum + dW3; end dW3avg = dW3sum / M; dW4avg = dW4sum / M; W3=W3+dW3avg; W4=W4+dW4avg; end end