Issue with Image sizes when doing transfer learning on Alexnet

Hi,

I get the following error:

Error using nnet.internal.cnn.MiniBatchDatasourceDispatcher>iCellTo4DArray (line 328)
Unexpected image size: All images must have the same size.

Error in nnet.internal.cnn.MiniBatchDatasourceDispatcher/readObservations (line 253)
                observations = iCellTo4DArray(X);

Error in nnet.internal.cnn.MiniBatchDatasourceDispatcher/readData (line 245)
            miniBatchData = this.readObservations(X);

Error in nnet.internal.cnn.MiniBatchDatasourceDispatcher/next (line 144)
                [miniBatchData, miniBatchResponse] = readData(this);

Error in SeriesNetwork/predict (line 337)
                        [X, ~, i] = dispatcher.next();

Error in SeriesNetwork/classify (line 559)
            scores = this.predict( X, varargin{:} );

When trying to retrain alexnet with my own imagesets. The directory contains images that is not in 227 by 227 size, so Im resizing them. It fails on testpreds = classify(XXnet,testImgs); I've tried the current code with the ReadFcn, but I've also tried a augmentedImageSource for testing. I cant get this to work.

Since I try to use the same code for resizing both training and testing images, I dont really understand this at all.

XXds = imageDatastore('ImageDir','IncludeSubfolders',true,'LabelSource','foldernames');

testImgs.ReadFcn=@(images_set)imresize(imread(images_set),[227 227]);
[trainImgs,testImgs] = splitEachLabel(XXds,0.6);

trainImgs.ReadFcn=@(trainImgs)imresize(imread(trainImgs),[227 227]);
XXTrain = augmentedImageSource([227 227], trainImgs)

testImgs.ReadFcn = @(testImgs)imresize(imread(testImgs),[227 227])
testImgsBatch = imresize(testImgs.read(), [227 227]);

numClasses = numel(categories(XXds.Labels));

net = alexnet;
layers = net.Layers;

layers(end-2) = fullyConnectedLayer(numClasses);
layers(end) = classificationLayer;

options = trainingOptions('sgdm','InitialLearnRate', 0.001,'MaxEpochs', 1);

[XXnet,info] = trainNetwork(XXTrain, layers, options);

testpreds = classify(XXnet,testImgs);

nnz(testpreds == testImgs.Labels)/numel(testpreds)

[XXconf,XXnames] = confusionmat(testImgs.Labels,testpreds);
heatmap(XXnames,XXnames,XXconf);

Can anyone see what Im doing wrong here?

Answer 1

[Sourabh]

Hey @Emil Alne

As per my understanding, AlexNet requires RGB images. Using “augmentedImageDatastore” for training ensures correct size but not using it for testing may be the reason why the size check fails during classify.

Kindly refer the following code snippet to fix this issue:

XXds = imageDatastore('ImageDir', ...
    'IncludeSubfolders', true, ...
    'LabelSource', 'foldernames');

% Split dataset
[trainImgs, testImgs] = splitEachLabel(XXds, 0.6);

% Use augmentedImageDatastore to resize and ensure RGB format
XXTrain = augmentedImageDatastore([227 227 3], trainImgs);
XXTest = augmentedImageDatastore([227 227 3], testImgs);

% Setup AlexNet for transfer learning
net = alexnet;
layers = net.Layers;
numClasses = numel(categories(XXds.Labels));
layers(end-2) = fullyConnectedLayer(numClasses);
layers(end) = classificationLayer;

% Training options
options = trainingOptions('sgdm', ...
    'InitialLearnRate', 0.001, ...
    'MaxEpochs', 1);

% Train
[XXnet, info] = trainNetwork(XXTrain, layers, options);

% Classify test images
testpreds = classify(XXnet, XXTest);

% Evaluate
accuracy = nnz(testpreds == testImgs.Labels) / numel(testpreds);
disp(['Accuracy: ' num2str(accuracy)]);

[XXconf, XXnames] = confusionmat(testImgs.Labels, testpreds);
heatmap(XXnames, XXnames, XXconf);

For more information on “augmentedImageDatastore” and “alexnet”, kindly refer the following MATLAB documentation:

• https://www.mathworks.com/help/deeplearning/ref/augmentedimagedatastore.html
• https://www.mathworks.com/help/deeplearning/ref/alexnet.html