将数字识别神经网络转换为定点并生成C代码

%加载数据Tr = csvread()“train.csv”， 1,0);%读取train.csvSub = csvread()“test.csv”， 1,0);%读取test.csv%准备数据N = size(tr, 1);数据集中样本数的%target = tr(:，1);%第一列是|标签|目标(目标== 0)= 10;%用“10”表示“0”Targetsd = dummyvar(目标);%将标签转换为虚拟变量输入= tr(:，2:end);%其余列是预测符输入=输入';转置输入%目标=目标';转置目标Targetsd = Targetsd ';%转置哑变量rng (1);再现性%C = cvpartition(n，“坚持”、n / 3);%保留数据集的1/3Xtrain = inputs(:， training(c));% 2/3的投入用于培训Ytrain = targetsd(:， training(c));%培训目标的2/3Xtest = inputs(:， test(c));% 1/3的输入用于测试Ytest = targets(test(c));%目标的1/3用于测试Ytestd = targetsd(:， test(c));% 1/3用于测试的虚拟变量%列车网络hiddenLayerSize = 100;net = patternnet(hiddenLayerSize);[net, tr] = train(net, Xtrain, Ytrain);视图(净);输出= net(Xtest);errors = gsubtract(Ytest，输出);performance = perform(net, Ytest, output);图中,plotperform (tr);

nnet.guis.closeAllViews ();

Sys_name = gensim(net，“名字”，“mTrainedNN”）;

模型=“fxpdemo_mnist_classification”;System_under_design =[模型/模式识别神经网络];Baseline_output =[模型' / yarr '];open_system(模型);

converter = DataTypeWorkflow.Converter(system_under_design);

converter.applySettingsFromShortcut (“使用双重覆盖收集范围”）;collect_ranges = converter.CurrentRunName;sim_out = converter.simulateSystem();

plotConfusionMatrix(sim_out, baseline_output, system_under_design，“转换前的分类率”）;

ps = DataTypeWorkflow.ProposalSettings;转换器。proposeDataTypes (collect_ranges、ps);

converter.applyDataTypes (collect_ranges);

converter.applySettingsFromShortcut ('具有指定数据类型的范围集合'）;sim_out = converter.simulateSystem();

plotConfusionMatrix(sim_out, baseline_output, system_under_design，“定点换算后的分类费率”）;

Block_path = [system_under_design .“图层1 / tansig”];p = FunctionApproximation.Problem(block_path);p. options . wordlength = 16;p.Options.BreakpointSpecification =“EvenPow2Spacing”;Solution = p.solve;solution.replaceWithApproximate;

Block_path = [system_under_design ./ 2层/ softmax /经验的];p = FunctionApproximation.Problem(block_path);p. options . wordlength = 16;p.Options.BreakpointSpecification =“EvenPow2Spacing”;

p. inputlowbounds = -40;p.InputUpperBounds = 0;Solution = p.solve;solution.replaceWithApproximate;

converter.applySettingsFromShortcut (converter.ShortcutsForSelectedSystem {2});sim_out = converter.simulateSystem;plotConfusionMatrix(sim_out, baseline_output, system_under_design，“功能替换后的分类率”）;