added load_images_MNIST() to utils

code/nnv/engine/utils/load_images_MNIST.m

@@ -0,0 +1,69 @@
+function [images, labels, db_image_nos] = load_images_MNIST(args)
+
+    arguments
+        args.database = "mnist"  
+        args.n               % number of images
+        args.matlabnet       % if a matlab neural network is supplied, only 
+            % the correctly classified images will be returned
+    end
+
+    database = args.database;
+    n = args.n;
+    matlabnet = args.matlabnet;
+
+    images = {};
+    labels = {};
+    db_image_nos = {};
+    if strcmp(database, "mnist")
+        if n > 10000
+            error('Maximum 10000 mnist images available.')
+        end
+        % Load data (no download necessary)
+        digitDatasetPath = fullfile(matlabroot,'toolbox','nnet','nndemos', ...
+            'nndatasets','DigitDataset');
+        % Images
+        imds = imageDatastore(digitDatasetPath, ...
+            'IncludeSubfolders',true,'LabelSource','foldernames');
+
+        numClasses = 10;
+        if mod(n, numClasses) ~= 0
+            error(['For MNIST, to have balanced dataset, number of images must be a multiple of ' num2str(numClasses)])
+        end
+        NPerClass = n/numClasses;
+
+        db_img_no = 1;
+        no_of_images_chosen_from_class = 0;
+        class_no = 1;
+        while class_no <= numClasses
+            % Load one image in dataset
+            [img, fileInfo] = readimage(imds, db_img_no);
+            img = single(img); % change precision
+            label = single(fileInfo.Label);
+
+            append_image = 1;
+            if ~isempty(matlabnet)
+                [~, pred] = max(predict(matlabnet, img));
+                if label ~= pred
+                    append_image = 0;
+                end
+            end
+
+            if append_image
+                images{end + 1} = img;
+                labels{end + 1} = label;
+                db_image_nos{end + 1} = db_img_no;
+                no_of_images_chosen_from_class = no_of_images_chosen_from_class + 1;
+                if no_of_images_chosen_from_class == NPerClass
+                    class_no = class_no + 1;
+                    no_of_images_chosen_from_class = 0;
+                    db_img_no = (class_no - 1)*1000;
+                end
+            end
+            db_img_no = db_img_no + 1;
+        end
+
+    else
+        error(['Unsupported database ' database])
+    end
+end
+

code/nnv/examples/Tutorial/NN/MNIST/weightPerturb/single_layer/conv_expt_any_layer.m

@@ -60,7 +60,7 @@
     otherwise
         error("Not supported yet.")
 end
-[images, labels] = load_images(database = database, ...
+[images, labels] = load_images_MNIST(database = database, ...
     n = expt.data.n_images, ...
     matlabnet = matlabnet);
 

code/nnv/tests/weightPerturb/test_conv_layer_perturbation.m

@@ -24,13 +24,8 @@
 reachOptions.numCores = 1;
 % reachOptions.device = 'gpu';
 reachOptions.device = 'cpu';
-% reachOptions.delete_old_sets = 1;
-reachOptions.free_mem_frac_for_LPs_in_verify_specification = 0.1;
-% reachOptions.dis_opt = 'display';
-% reachOptions.disp_intersection_result = 1;
-% reachOptions.debug = 1;
 
-[images, labels] = load_images(database = "mnist", ...
+[images, labels] = load_images_MNIST(database = "mnist", ...
     n = 10, ...
     matlabnet = matlabnet);
 img = images{1};