I understand that you're trying to use VGG-16 as a feature extractor for your images and then use an SVM classifier to classify those images.
Reading the error message, it says X and Y do not have the same number of observations. To use the fitcecoc function, the features and the labels should have the same number of observations. If there are m images, then X should be 
, where n is the number of features, and Y should be 
.
, where n is the number of features, and Y should be .In the code above, X (X_train) is a 
row vector and Y (Y_train) is a 
row vector. Please consider the explanation below to understand why this is so.
row vector and Y (Y_train) is a row vector. Please consider the explanation below to understand why this is so.Executing the following code:
healthy_features = activations(net, healthy_images, 'fc7');
The output is 
since the fc7 layer in VGG-16 has 
output units and there are 
healthy images. The first two dimensions are supposed to represent the height and width respectively but since this is a fully-connected layer, the height and width are 1. The other feature vectors have similar dimensions and differ only in the last dimension since there are different number of images. This behavior of the activations function is documented here: https://www.mathworks.com/help/deeplearning/ref/seriesnetwork.activations.html#mw_b65e2845-af4a-42ab-ae94-cbd612c5324b.
since the fc7 layer in VGG-16 has output units and there are healthy images. The first two dimensions are supposed to represent the height and width respectively but since this is a fully-connected layer, the height and width are 1. The other feature vectors have similar dimensions and differ only in the last dimension since there are different number of images. This behavior of the activations function is documented here: https://www.mathworks.com/help/deeplearning/ref/seriesnetwork.activations.html#mw_b65e2845-af4a-42ab-ae94-cbd612c5324b.Now, the following code:
min_size = min([size(healthy_features, 1), size(mild_DR_features, 1), size(moderate_DR_features, 1), size(proliferate_DR_features, 1), size(severe_DR_features, 1)]);
Sets min_size to 1 since the first dimension of each feature vector is 1. Then, the code:
healthy_features = healthy_features(1:min_size, :);
Flattens the feature vector to 
(
). Similarly, the other feature vectors are flattened. The code:
(). Similarly, the other feature vectors are flattened. The code:X_train = [healthy_features, mild_DR_features, moderate_DR_features, proliferate_DR_features, severe_DR_features];
Concatenates all the flattened vectors into a single row vector, resulting in the row vector. Also, since min_size is 1, the code:
row vector. Also, since min_size is 1, the code:Y_train = [ones(min_size, 1),2*ones(min_size, 1),3*ones(min_size, 1), 4*ones(min_size, 1), 5*ones(min_size, 1)];
Creates a simple vector with the value [1,2,3,4,5], resulting in the row vector.
row vector.Manually creating a datastore for each label, extracting the features and concatenating them can be avoided by:
- Using a single datastore for the dataset by specifying the IncludeSubfolders option and setting LabelSource to foldernames when creating the datastore. This will ensure that images from each subfolder are part of the datastore and that each image is assigned a label according to the subfolder it is in. This will also allow you to use the Labels property to create Y_train.
- Setting the OutputAs option to rows to ensure that the activations function outputs the correct dimension. This will ensure that the output from the function is .
Assuming that the root data folder is colored_images/, the code below should fix the issue (I am omitting the plotting and evaluation code for brevity):
imds = imageDatastore("colored_images\", "IncludeSubfolders", true, "LabelSource", "foldernames");
% Plotting code
net = vgg16;
X_train = activations(net, imds, "fc7", "OutputAs", "rows");
Y_train = imds.Labels;
svm_model = fitcecoc(X_train, Y_train);
% Evaludation code
Hope this helps!
