downloadFolder = matlab.internal.examples.downloadSupportFile (“音频”,“AirCompressorDataset / AirCompressorDataset.zip”);dataFolder = tempdir;unzip(downloadFolder,dataFolder) dataset = fullfile(dataFolder，“AirCompressorDataset”);

广告= audioDatastore(数据集,IncludeSubfolders = true, LabelSource =“foldernames”);rng默认的[adsTrain, adsValidation] = splitEachLabel(广告,0.8);

uniqueLabels =独特(adsTrain.Labels);tblTrain = countEachLabel (adsTrain);tblValidation = countEachLabel (adsValidation);H =酒吧(uniqueLabels, [tblTrain。计数,tblValidation。数),“堆叠”);传奇(H, [“训练集”,“验证”),位置=“NorthEastOutside”)

t = (0:5e4-1) / 16 e3;tiledlayout (4,2 TileSpacing =“紧凑”填充=“紧凑”)为n = 1:numel(uniqueLabels) idx = find(adsTrain.Labels==uniqueLabels(n));[x, fs] = audioread (adsTrain.Files {idx (randperm(元素个数(idx), 1))});nexttile plotHandle = plot(t,x);如果N == 7 ||“秒”);其他的甘氨胆酸集(xtick = [])结束标题(string (uniqueLabels (n)));结束

afe = audioFeatureExtractor (SampleRate = fs,…FFTLength = 4096,…窗口=损害(圆(fs * 0.2),“周期”),…OverlapLength =圆(fs * 0.195),…melSpectrum = true);setExtractorParameters (afe“melSpectrum”NumBands = 128);featureVector =提取(afe x);[numHops, numFeatures numChannels] =大小(featureVector)

抵消=每股收益;adsTrainCPU =变换(adsTrain @ (x) log10(提取(afe x) +偏移量));mixerCPU =混合物(adsTrainCPU);adsTrainCPU =变换(adsTrainCPU @ (x,信息)混合(mixerCPU x,信息),IncludeInfo = true);adsTrainGPU =变换(adsTrain @gpuArray);adsTrainGPU =变换(adsTrainGPU @ (x) log10(提取(afe x) +偏移量));mixerGPU =混合物(adsTrainGPU);adsTrainGPU =变换(adsTrainGPU @ (x,信息)混合(mixerGPU x,信息),IncludeInfo = true);

adsValidationGPU =变换(adsValidation @gpuArray);adsValidationGPU =变换(adsValidationGPU @ (x) {log10(提取(afe x) +偏移量)});adsValidationGPU =结合(adsValidationGPU arrayDatastore (adsValidation.Labels));adsValidationCPU =变换(adsValidation @ (x) {log10(提取(afe x) +偏移量)});adsValidationCPU =结合(adsValidationCPU arrayDatastore (adsValidation.Labels));

抽搐为ii = 1:numel(adsTrain.Files) x = read(adsTrainCPU);结束cpuPipeline = toc;重置(adsTrainCPU)抽搐为ii = 1:numel(adsTrain.Files) x = read(adsTrainGPU);结束等待(gpuDevice)%确保所有计算都已完成gpuPipeline = toc;重置(adsTrainGPU) disp ([“读取、提取和扩充训练集(CPU):”+ cpuPipeline +“秒”;…“读取、提取和扩充列车集(GPU):”+ gpuPipeline +“秒”;…“Speedup (CPU时间)/(GPU时间):”+ cpuPipeline / gpuPipeline]);

x =阅读(广告);提取(afe x);%在计时循环之外产生初始化成本抽搐为ii = 1:numel(adsTrain.Files) features = log10(extract(afe,x)+offset);结束cpuFeatureExtraction = toc;x = gpuArray (x);%在计时循环之外产生初始化成本提取(afe x);抽搐为ii = 1:numel(adsTrain.Files) features = log10(extract(afe,x)+offset);结束等待(gpuDevice)%确保所有计算都已完成gpuFeatureExtraction = toc;disp ([“从训练集(CPU)中提取特征:”+ cpuFeatureExtraction +“秒”;…“从训练集(GPU)中提取特征:”+ gpuFeatureExtraction +“秒”;…“Speedup (CPU时间)/(GPU时间):”+ cpuFeatureExtraction / gpuFeatureExtraction]);

numClasses =元素个数(类别(adsTrain.Labels));图象尺寸= [numHops, afe.FeatureVectorLength];layers = [imageInputLayer(imageSize,Normalization= .“没有”现年48岁的)convolution2dLayer(3填充=“相同”) batchNormalizationLayer reluLayer maxPooling2dLayer([numHops,1]) fullyConnectedLayer(numClasses) softmaxLayer classificationLayer(Classes=categories(adsTrain.Labels));];

miniBatchSize = 128;选择= trainingOptions (“亚当”,…洗牌=“every-epoch”,…MaxEpochs = 40,…LearnRateSchedule =“分段”,…LearnRateDropPeriod = 15,…LearnRateDropFactor = 0.2,…MiniBatchSize = MiniBatchSize,…情节=“训练进步”,…Verbose = false,…ValidationData = adsValidationCPU,…ValidationFrequency =装天花板(元素个数(adsTrain.Files) / miniBatchSize),…ExecutionEnvironment =“multi-gpu”);

tic net = trainNetwork(adsTrainCPU,layers,options);

cpuTrainTime = toc;

选项。ValidationData = adsValidationGPU;tic net = trainNetwork(adsTrainGPU,layers,options);

gpuTrainTime = toc;

disp (["训练时间(CPU): "+ cpuTrainTime +“秒”;“训练时间(GPU):”+ gpuTrainTime +“秒”;“Speedup (CPU时间)/(GPU时间):”+ cpuTrainTime / gpuTrainTime])

signalToClassify =阅读(广告);gpuFeatureExtraction = gputimeit(@()预测(净,log10(提取(afe, gpuArray (signalToClassify)) +偏移量)));cpuFeatureExtraction = gputimeit(@()预测(净,log10(提取(afe (signalToClassify)) +偏移量)));disp ([“3秒数据的预测时间(CPU上的特征提取):”e3 + + cpuFeatureExtraction * 1“女士”;…“3 s数据的预测时间(GPU上的特征提取):”e3 + + gpuFeatureExtraction * 1“女士”;…“Speedup (CPU时间)/(GPU时间):”+ cpuFeatureExtraction / gpuFeatureExtraction])

adsValidationGPU =变换(adsValidation @ (x) gpuArray (x));adsValidationGPU =变换(adsValidationGPU @ (x) {log10(提取(afe x) +偏移量)});adsValidationCPU =变换(adsValidation @ (x) {log10(提取(afe x) +偏移量)});gpuFeatureExtraction = gputimeit(@()预测(净、adsValidationGPU ExecutionEnvironment =“multi-gpu”));cpuFeatureExtraction = gputimeit(@()预测(净、adsValidationCPU ExecutionEnvironment =“multi-gpu”));disp ([“验证集预测时间(CPU上的特征提取):”+ cpuFeatureExtraction +“秒”;“验证集的预测时间(GPU上的特征提取):”+ gpuFeatureExtraction +“秒”;“Speedup (CPU时间)/(GPU时间):”+ cpuFeatureExtraction / gpuFeatureExtraction])

类型混合