Datasets
Models
Diff of
/steps/step_C_varpcanet.m
[000000]
..
[1422d3]
Switch to side-by-side view
--- a +++ b/steps/step_C_varpcanet.m @@ -0,0 +1,548 @@ +function [] = step_C_varpcanet(dirTest, dirUtilities, ext, numCoresFeatExtr, numCoresKnn, fidLogs, logS, savefile, plotFigures) + +%-------------------------------------- +%General parameters +stepPrint = 100; +%PCA Params +run('./params/paramsPCATuning.m'); + + +%-------------------------------------- +%Dir DBs +dbname_All = { ... + 'ALL_IDB' + }; +dbname_part_All = { ... + 'ALL_IDB2' + }; +dbname_ROI_All = { ... + 'ROI_256' + }; + +%nets +net_name{1} = 'AlexNet'; +net_name{2} = 'VGG16'; +net_name{3} = 'VGG19'; +net_name{4} = 'ResNet18'; +net_name{5} = 'ResNet50'; +net_name{6} = 'ResNet101'; +net_name{7} = 'DenseNet201'; + + +%-------------------------------------- +colorS_init = 1; +colorS_tune = 1; +colorS_test = 1; +%-------------------------------------- + +% processDummyDirs(); + + + +%-------------------------------------- +%Loop on dbs +for db = 1 : numel(dbname_All) + % for db = 2 + + %Close + close all + pause(0.2); + + %DB selection + dbname = dbname_All{db}; + dbnamePart = dbname_part_All{db}; + ROI = dbname_ROI_All{db}; + dirDB_wROI = [dirTest dbname '/' dbnamePart '/' ROI '/']; + %dirDB_noROI = [dirWorkspace dbname '/' dbnamePart '/' ]; + + + %-------------------------------------- + %loop on nets + for n = 1 : numel(net_name) + + %switch net + switch n + + case 1 + net = alexnet; + layer = 'fc6'; + conv_layer = 'conv5'; + + case 2 + net = vgg16; + layer = 'fc6'; + conv_layer = 'conv5_3'; + + case 3 + net = vgg19; + layer = 'fc6'; + conv_layer = 'conv5_4'; + + case 4 + net = resnet18; + layer = 'fc1000'; + conv_layer = 'res5b_relu'; + + case 5 + net = resnet50; + layer = 'fc1000'; + conv_layer = 'res5c_branch2c'; + + case 6 + net = resnet101; + layer = 'fc1000'; + conv_layer = 'res5c_branch2c'; + + case 7 + net = densenet201; + layer = 'fc1000'; + conv_layer = 'conv5_block32_2_conv'; + + end %switch + + + %-------------------------------------- + %Folder creation + %RESULTS: dirs net + dirResults = ['./Results/' dbname '/' dbnamePart '/' net_name{n} '/']; + mkdir_pers(dirResults, savefile); + %RESULTS: log file + timeStampRaw = datestr(datetime); + timeStamp = strrep(timeStampRaw, ':', '-'); + if savefile && logS + logFile = [dirResults dbname '_log_' timeStamp '.txt']; + fidLog = fopen(logFile, 'w'); + fidLogs{2} = fidLog; + end %if savefile && log + + + %-------------------------------------- + %Display + fprintf_pers(fidLogs, '\n'); + fprintf_pers(fidLogs, '---------------\n'); + fprintf_pers(fidLogs, 'ALL-Unsharpen\n'); + fprintf_pers(fidLogs, [dbname '\n']); + fprintf_pers(fidLogs, [dbnamePart '\n']); + fprintf_pers(fidLogs, '---------------\n'); + fprintf_pers(fidLogs, '\n'); + + + %-------------------------------------- + %display + fprintf_pers(fidLogs, '---------------\n'); + fprintf_pers(fidLogs, ['Net: ' net_name{n} '\n']); + fprintf_pers(fidLogs, '---------------\n'); + fprintf_pers(fidLogs, '\n') + + + %-------------------------------------- + %DB processing + %Extract samples + files = dir([dirDB_wROI '*.' ext]); + + %Compute labels + [problem, labels, numImagesAll] = computeLabels(dirDB_wROI, files); + + + %-------------------------------------- + %Display + fprintf_pers(fidLogs, 'Extracting samples...\n'); + fprintf_pers(fidLogs, ['\t' num2str(numImagesAll) ' images in total\n']); + fprintf_pers(fidLogs, '\n'); + + + %-------------------------------------- + %LOOP ON ITERATIONS + %Init + accuracy_knnAll = zeros(param.numIterations, 1); + cmc = cell(param.numIterations, 1); + + %-------------------------------------- + %Compute random fold indexes + %if outside iteration loop, random fold once (es. 10-fold) + %[allIndexes, cvIndices] = computeAllIndexesFold(numImagesAll, labels, param); + + %Loop + for r = 1 : param.numIterations + + + %-------------------------------------- + %Display + fprintf_pers(fidLogs, ['Iteration N. ' num2str(r) '\n']); + + + %-------------------------------------- + %File save info + fileSaveTest_iter = [dirResults '/results_iter_' num2str(r) '.mat']; + + + %-------------------------------------- + %Compute random fold indexes + %--10-fold + %[indImagesTrain, indImagesTest, numImagesTrain, numImagesTest] = computeIndexesFold(cvIndices, r); + %--2-fold if inside iteration loop, random fold each iteration (repeated 2-fold) + [allIndexes, cvIndices] = computeAllIndexesFold(numImagesAll, labels, param); + [indImagesTrain, indImagesTest, numImagesTrain, numImagesTest] = computeIndexesFold(cvIndices, randi(2, 1)); + %Corresponding labels + TrnLabels = labels(indImagesTrain); + TestLabels = labels(indImagesTest); + + + %-------------------------------------- + %Display output number of images + fprintf_pers(fidLogs, ['\t' num2str(numImagesTrain) ' images are chosen for training\n']); + fprintf_pers(fidLogs, ['\t' num2str(numImagesTest) ' images are chosen for testing\n']); + + + + %%%%%%%%%%%%%% TRAINING %%%%%%%%%%%%% + start_pool(numCoresFeatExtr); + + %-------------------------------------- + fprintf_pers(fidLogs, '\tTraining... \n') + + + %-------------------------------------- + %Load images for training + fprintf_pers(fidLogs, '\t\tLoading images for training... \n') + [imagesCellTrain, filenameTrn, ~] = loadImages(files, dirDB_wROI, allIndexes, indImagesTrain, numImagesTrain, param, colorS_init, 100, dirUtilities, 0); + imagesCellTrain = adjustFormat(imagesCellTrain); + + %norm + %[imagesCellTrain, dd1, dd2] = computeNorm(imagesCellTrain); + + %find th_focus + fprintf_pers(fidLogs, '\tComputing th_focus init... \n') + + %%%%%% QUI %%%%%% + th_focus_init = find_th_focus(imagesCellTrain, TrnLabels, [128 128], dirUtilities, fidLogs); + %th_focus_init = 7.3; + + %puliamo + clear imagesCellTrain + + %tune + %th_focus_start = round((th_focus_init - th_focus_init*percC/100)*10)/10; + %th_focus_end = round((th_focus_init + th_focus_init*percC/100)*10)/10; + th_focus_start = round((th_focus_init - 0.5)*10)/10; + th_focus_end = round((th_focus_init + 0.5)*10)/10; + %th_focus_start = th_focus_init; + %th_focus_end = th_focus_init; + + %init + accuracy_knnALLFOCUS = []; + + + %-------------------------------------- + %tuning th_focus + fprintf_pers(fidLogs, '\tTuning th_focus... \n') + + %loop on th_focus + allfocuses = th_focus_start : 0.1 : th_focus_end; + %%%%%% QUI %%%%%% + for th_focus = allfocuses + %for th_focus = th_focus_start + + %Display + fprintf_pers(fidLogs, '\n') + fprintf_pers(fidLogs, ['\t\tth_focus: ' num2str(th_focus) '\n']); + + + fprintf_pers(fidLogs, '\t\tLoading images... \n') + [imagesCellTrain, filenameTrn, ~] = loadImages(files, dirDB_wROI, allIndexes, indImagesTrain, numImagesTrain, param, colorS_tune, th_focus, dirUtilities, 0); + imagesCellTrain = adjustFormatForPCANet(imagesCellTrain, [128, 128]); + + %-------------------------------------- + %PCANet Training + %1 layer: PCA filters + [V, PCANet] = trainPCANet(imagesCellTrain, PCANet, fidLogs, param, numCoresFeatExtr); + %Feature extraction + fprintf_pers(fidLogs, '\t\tFeature extraction... \n') + [ftrain_all, numFeaturesTrain] = featExtrGaborAdapt(imagesCellTrain, V, PCANet, [], param, numImagesTrain, stepPrint); + + %size + sizeTrain = size(ftrain_all, 2); + + %-------------------------------------- + %performance + fprintf_pers(fidLogs, '\t\tClassification - original... \n') + errorStruct_original_temp = computeClassificationPerformance(numFeaturesTrain, sizeTrain, ftrain_all, TrnLabels, stepPrint, numCoresKnn, fidLogs, param); + + %compute cmc + [cmc_original_temp, cmc_sum_original_temp] = computeCMC(errorStruct_original_temp.distMatrixTest, TrnLabels, ['cmc original temp iteration ' num2str(r)], 0); + + %Error metrics + fprintf_pers(fidLogs, ['\t\tTraining accuracy (perc. of correctly classified samples, at iteration n. ' num2str(r) '): %s%%\n'], num2str(errorStruct_original_temp.accuracy_knn*100)); + fprintf_pers(fidLogs, ['\t\tAUC of CMC (at iteration n. ' num2str(r) '): %s\n'], num2str(cmc_sum_original_temp)); + + + %assign + accuracy_knnALLFOCUS = [accuracy_knnALLFOCUS errorStruct_original_temp.accuracy_knn]; + %accuracy_knnALLFOCUS = [accuracy_knnALLFOCUS cmc_sum_original_temp]; %maximize AUC of CMC + + + end %th_focus + + %Puliamo + clear imagesCellTrain ftrain_all + + + + + + %%%%%%%%%%%%%% APPLY BEST FOCUS %%%%%%%%%%%%% + fprintf_pers(fidLogs, '\n') + fprintf_pers(fidLogs, '\tApply best focus\n') + %best focusc + %[maxAcc, i_best_focus] = max(accuracy_knnALLFOCUS); + %i_best_focus = i_best_focus(1); + [sortA, isort] = sort(accuracy_knnALLFOCUS); + i_best_focus = isort(end); %highest th_focus that gives the best result + best_th_focus = allfocuses(i_best_focus); + fprintf_pers(fidLogs, ['\t\tBest focus: ' num2str(best_th_focus) ' \n']); + + %training data + fprintf_pers(fidLogs, '\t\tLoading training images - original... \n') + [imagesCellTrain_original, ~, ~] = loadImages(files, dirDB_wROI, allIndexes, indImagesTrain, numImagesTrain, param, colorS_test, 100, dirUtilities, 0); + imagesCellTrain_original = adjustFormat(imagesCellTrain_original); + %norm + %[imagesCellTrain_original, meanA_original, stdA_original] = computeNorm(imagesCellTrain_original); + fprintf_pers(fidLogs, '\t\tLoading training images - unsharpened... \n') + [imagesCellTrain_unsharp, ~, ~] = loadImages(files, dirDB_wROI, allIndexes, indImagesTrain, numImagesTrain, param, colorS_test, best_th_focus, dirUtilities, 0); + imagesCellTrain_unsharp = adjustFormat(imagesCellTrain_unsharp); + %norm + %[imagesCellTrain_unsharp, meanA_unsharp, stdA_unsharp] = computeNorm(imagesCellTrain_unsharp); + + %testing data + fprintf_pers(fidLogs, '\t\tLoading testing images - original... \n') + [imagesCellTest_original, ~, ~] = loadImages(files, dirDB_wROI, allIndexes, indImagesTest, numImagesTest, param, colorS_test, 100, dirUtilities, 0); + [imagesCellTest_original, meanAll_test_original] = adjustFormat(imagesCellTest_original); + %imagesCellTest_original = applyNorm(imagesCellTest_original, meanA_original, stdA_original); + fprintf_pers(fidLogs, '\t\tLoading testing images - unsharpened... \n') + [imagesCellTest_unsharp, filenameTest, indexes_test_imUnsharpened] = loadImages(files, dirDB_wROI, allIndexes, indImagesTest, numImagesTest, param, colorS_test, best_th_focus, dirUtilities, 0); + [imagesCellTest_unsharp, meanAll_test_unsharp] = adjustFormat(imagesCellTest_unsharp); + %imagesCellTest_unsharpened = applyNorm(imagesCellTest_unsharpened, meanA_unsharp, stdA_unsharp); + + + + %%%%%%%%%%%%%% TESTING - PRE-TRAINED CNNs %%%%%%%%%%%%% + %pre-trained + fprintf_pers(fidLogs, '\t\tPretrained CNNs... \n') + + + %Feature extraction - ORIGINAL + fprintf_pers(fidLogs, '\t\t\tFeature extraction - original... \n') + ftrain_all_original = feature_extraction_cnn(imagesCellTrain_original, net, layer, colorS_test); + ftest_all_original = feature_extraction_cnn(imagesCellTest_original, net, layer, colorS_test); + + %Feature extraction - UNSHARP + fprintf_pers(fidLogs, '\t\t\tFeature extraction - unsharp... \n') + ftrain_all_unsharp = feature_extraction_cnn(imagesCellTrain_unsharp, net, layer, colorS_test); + ftest_all_unsharp = feature_extraction_cnn(imagesCellTest_unsharp, net, layer, colorS_test); + + %size + numFeatures = size(ftest_all_unsharp, 1); + sizeTest = size(ftest_all_unsharp, 2); + + + %-------------------------------------- + %Classification performance + %Original + %fprintf_pers(fidLogs, '\t\t\tClassification - original... \n') + %errorStruct_pretrained_original(r) = computeClassificationPerformance(numFeatures, sizeTest, ftest_all_original, TestLabels, stepPrint, numCoresKnn, fidLogs, param); + errorStruct_pretrained_original(r) = computeClassificationPerformanceTrainTest(numFeatures, sizeTest, ftrain_all_original, ftest_all_original, TrnLabels, TestLabels, stepPrint, numCoresKnn, fidLogs, param); + %Unsharp + %fprintf_pers(fidLogs, '\t\t\tClassification - unsharp... \n') + %errorStruct_pretrained_unsharp(r) = computeClassificationPerformance(numFeatures, sizeTest, ftest_all_unsharp, TestLabels, stepPrint, numCoresKnn, fidLogs, param); + errorStruct_pretrained_unsharp(r) = computeClassificationPerformanceTrainTest(numFeatures, sizeTest, ftrain_all_unsharp, ftest_all_unsharp, TrnLabels, TestLabels, stepPrint, numCoresKnn, fidLogs, param); + + %puliamo + clear ftest_all_original ftest_all_unsharp + + %compute cmc + [cmc1, cmc_sum1] = computeCMC_trainTest(errorStruct_pretrained_original(r).distMatrixTest, TestLabels, ['cmc original iteration ' num2str(r)], plotFigures); + [cmc2, cmc_sum2] = computeCMC_trainTest(errorStruct_pretrained_unsharp(r).distMatrixTest, TestLabels, ['cmc unsharp iteration ' num2str(r)], plotFigures); + cmc_original{r} = cmc1; + cmc_sum_original{r} = cmc_sum1; + cmc_unsharp{r} = cmc2; + cmc_sum_unsharp{r} = cmc_sum2; + errorStruct_pretrained_original(r).rank5 = cmc1(5); + errorStruct_pretrained_unsharp(r).rank5 = cmc2(5); + + %Display + fprintf_pers(fidLogs, '\n') + fprintf_pers(fidLogs, ['\tPretrained - Accuracy original (at iteration n. ' num2str(r) '): %s%%\n'], num2str(errorStruct_pretrained_original(r).accuracy_knn*100)); + fprintf_pers(fidLogs, ['\tPretrained - Accuracy unsharp (at iteration n. ' num2str(r) '): %s%%\n'], num2str(errorStruct_pretrained_unsharp(r).accuracy_knn*100)); + fprintf_pers(fidLogs, ['\tPretrained - Rank 5 accuracy original (at iteration n. ' num2str(r) '): %s%%\n'], num2str(cmc1(5)*100)); + fprintf_pers(fidLogs, ['\tPretrained - Rank 5 accuracy unsharp (at iteration n. ' num2str(r) '): %s%%\n'], num2str(cmc2(5)*100)); + + pause(0.1) + + + %%%%%%%%%%%%%% TESTING - FINE TUNING CNNs %%%%%%%%%%%%% + %pre-trained + fprintf_pers(fidLogs, '\n') + fprintf_pers(fidLogs, '\tFine tuning CNNs... \n') + %fprintf_pers(fidLogs, '\t\t\tTraining... \n') + + pixelRange = [-30 30]; + rotRange = [-180 180]; + + imageAugmenter = imageDataAugmenter( ... + 'RandXReflection', true, ... + 'RandYReflection', true, ... + 'RandRotation', rotRange); + %'RandXTranslation', pixelRange, ... + %'RandYTranslation', pixelRange ... + + + numClasses = numel(unique(labels)); + inputSize = net.Layers(1).InputSize; + % layersTransfer = net.Layers(1:end-3); + % layers = [ + % layersTransfer + % fullyConnectedLayer(numClasses, 'WeightLearnRateFactor', 20, 'BiasLearnRateFactor', 20) + % softmaxLayer + % classificationLayer]; + + %change last layers + lgraph = replaceLayers(net, numClasses); + + %options + options = trainingOptions('sgdm', ... + 'MiniBatchSize', 20, ... %128 20 + 'MaxEpochs', 100, ... + 'InitialLearnRate', 1e-4, ... + 'Shuffle', 'every-epoch', ... 'never' + 'ValidationFrequency', 3, ... + 'Verbose', false, ... + 'Plots', 'none'); % 'training-progress' + + fprintf_pers(fidLogs, '\t\tTraining original... \n') + %netTransfer_original = fineTuneCNN(imagesCellTrain_original, TrnLabels, './dummy_train_original/', inputSize, imageAugmenter, layers, options); + netTransfer_original = fineTuneCNN(imagesCellTrain_original, TrnLabels, './dummy_train_original/', inputSize, imageAugmenter, lgraph, options); + + fprintf_pers(fidLogs, '\t\tTraining unsharp... \n') + %netTransfer_unsharp = fineTuneCNN(imagesCellTrain_unsharp, TrnLabels, './dummy_train_unsharp/', inputSize, imageAugmenter, layers, options); + netTransfer_unsharp = fineTuneCNN(imagesCellTrain_unsharp, TrnLabels, './dummy_train_unsharp/', inputSize, imageAugmenter, lgraph, options); + + % + %fprintf_pers(fidLogs, '\t\t\tTesting... \n'); + + %cm + fprintf_pers(fidLogs, '\t\tTesting original... \n') + errorStruct_finetune_original(r) = computeClassPerformanceFineTuneCNN(imagesCellTest_original, TestLabels, './dummy_test_original/', inputSize, netTransfer_original, fidLogs); + + fprintf_pers(fidLogs, '\t\tTesting unsharp... \n') + errorStruct_finetune_unsharp(r) = computeClassPerformanceFineTuneCNN(imagesCellTest_unsharp, TestLabels, './dummy_test_unsharp/', inputSize, netTransfer_unsharp, fidLogs); + + %Display + fprintf_pers(fidLogs, ['\tFine tuning - Accuracy original (at iteration n. ' num2str(r) '): %s%%\n'], num2str(errorStruct_finetune_original(r).accuracy_knn*100)); + fprintf_pers(fidLogs, ['\tFine tuning - Accuracy unsharp (at iteration n. ' num2str(r) '): %s%%\n'], num2str(errorStruct_finetune_unsharp(r).accuracy_knn*100)); + + + %-------------------------------------- + %Save + if savefile + save(fileSaveTest_iter, 'errorStruct_pretrained_original', 'errorStruct_pretrained_unsharp', 'errorStruct_finetune_original', 'errorStruct_finetune_unsharp', 'cmc_original', 'cmc_unsharp'); + end %if savefile + + + %-------------------------------------- + %Display progress + fprintf_pers(fidLogs, '\n'); + + + %GRAD-CAM + fprintf_pers(fidLogs, 'Grad-CAM\n'); + fprintf_pers(fidLogs, '\n'); + dirGcam = [dirResults 'gcam_iter_' num2str(r) '/']; + mkdir_pers(dirGcam, savefile); + [imagesCellTest_original, ~, ~] = loadImages(files, dirDB_wROI, allIndexes, indImagesTest, numImagesTest, param, colorS_test, 100, dirUtilities, 0); + [imagesCellTest_unsharp, ~, ~] = loadImages(files, dirDB_wROI, allIndexes, indImagesTest, numImagesTest, param, colorS_test, best_th_focus, dirUtilities, 0); + computeGradCam(imagesCellTest_original, imagesCellTest_unsharp, meanAll_test_original, meanAll_test_unsharp, indexes_test_imUnsharpened, ... + netTransfer_original, netTransfer_unsharp, conv_layer, inputSize, filenameTest, TestLabels, dirGcam); + %computeGradCam2(imagesCellTest_original, imagesCellTest_unsharp, meanAll_test_original, meanAll_test_unsharp, indexes_test_imUnsharpened, ... + %netTransfer_original, netTransfer_unsharp, conv_layer, inputSize, filenameTest, TestLabels, dirGcam); + + + + + %Puliamo + clear imagesCellTrain_original imagesCellTrain_unsharpenend imagesCellTest_original imagesCellTest_unsharpened netTransfer_original netTransfer_unsharp layers + + + end %for r = 1 : param.numIterations + + + close all + pause(0.1) + + %display + fprintf_pers(fidLogs, '\n'); + + + + %-------------------------------------- + %Average classification performance + %PRETRAINED + %Error metrics + fprintf_pers(fidLogs, '\n'); + %original + fprintf_pers(fidLogs, 'Pretrained - Original\n') + stampaErrors(errorStruct_pretrained_original, fidLogs); + fprintf_pers(fidLogs, '\tRank 5 accuracy (mean; std): %s%%; %s%% \n', num2str(mean([errorStruct_pretrained_original.rank5])*100), num2str(std([errorStruct_pretrained_original.rank5])*100)); + %unsharp + fprintf_pers(fidLogs, 'Pretrained - Unsharp\n') + stampaErrors(errorStruct_pretrained_unsharp, fidLogs); + fprintf_pers(fidLogs, '\tRank 5 accuracy (mean; std): %s%%; %s%% \n', num2str(mean([errorStruct_pretrained_unsharp.rank5])*100), num2str(std([errorStruct_pretrained_unsharp.rank5])*100)); + + %FINE TUNING + %Error metrics + fprintf_pers(fidLogs, '\n'); + %original + fprintf_pers(fidLogs, 'Fine tuning - Original\n') + stampaErrors(errorStruct_finetune_original, fidLogs); + %unsharp + fprintf_pers(fidLogs, 'Fine tuning - Unsharp\n') + stampaErrors(errorStruct_finetune_unsharp, fidLogs); + + + + + %-------------------------------------- + %Average CMC + %original + stampaAvgCMC(cmc_original, 'original', dirResults, savefile, plotFigures); + %unsharp + stampaAvgCMC(cmc_unsharp, 'unsharp', dirResults, savefile, plotFigures); + + + + + %-------------------------------------- + %Display progress + fprintf_pers(fidLogs, '\n'); + + + %-------------------------------------- + %Close file log + if savefile && logS + fclose(fidLog); + end %if savefile && log + % delete(gcp('nocreate')); + fclose('all'); + + + %close + close all + pause(0.1) + + + end %for n + + +end %for db + + +