How to change output classes for image classification in alexnet?

Hello everyone

I'm trying to implement the following example:

https://www.mathworks.com/help/deeplearning/ug/transfer-learning-using-pretrained-network.html

but I got this error:

Error using trainNetwork (line 170)

Invalid training data. The output size (28224) of the last layer does not match the number of classes (24).

and this is my code:

tic
clear;
clc;
outputFolder = fullfile('Spilt_Dataset');
rootFolder = fullfile(outputFolder , 'Categories');
categories = {'front_or_left' , 'front_or_right','hump' , 'left_turn','narrow_from_left' , 'narrows_from_right',...
    'no_horron' , 'no_parking','no_u_turn' , 'overtaking_is_forbidden','parking' , 'pedestrian_crossing',...
    'right_or_left' , 'right_turn','rotor' , 'slow','speed_30' , 'speed_40',...
    'speed_50','speed_60','speed_80','speed_100','stop','u_turn'};
imds = imageDatastore(fullfile(rootFolder,categories),'LabelSource','Foldernames');
tbl = countEachLabel(imds);
minSetCount = min(tbl{:,2});
imds = splitEachLabel(imds, minSetCount, 'randomize');
countEachLabel(imds); 

front_or_left = find(imds.Labels == 'front_or_left',1);
front_or_right = find(imds.Labels == 'front_or_right',1);
hump = find(imds.Labels == 'hump',1);
left_turn = find(imds.Labels == 'left_turn',1);
narrow_from_left = find(imds.Labels == 'narrow_from_left',1);
narrows_from_right = find(imds.Labels == 'narrows_from_right',1);
no_horron = find(imds.Labels == 'no_horron',1);
no_parking = find(imds.Labels == 'no_parking',1);
no_u_turn = find(imds.Labels == 'no_u_turn',1);
overtaking_is_forbidden = find(imds.Labels == 'overtaking_is_forbidden',1);
parking = find(imds.Labels == 'parking',1);
pedestrian_crossing = find(imds.Labels == 'pedestrian_crossing',1);
right_or_left = find(imds.Labels == 'right_or_left',1);
right_turn = find(imds.Labels == 'right_turn',1);
rotor = find(imds.Labels == 'rotor',1);
slow = find(imds.Labels == 'slow',1);
speed_30 = find(imds.Labels == 'speed_30',1);
speed_40 = find(imds.Labels == 'speed_40',1);
speed_50 = find(imds.Labels == 'speed_50',1);
speed_60 = find(imds.Labels == 'speed_60',1);
speed_80 = find(imds.Labels == 'speed_80',1);
speed_100 = find(imds.Labels == 'speed_100',1);
stop = find(imds.Labels == 'stop',1);
u_turn = find(imds.Labels == 'u_turn',1);

[imdsTrain,imdsValidation] = splitEachLabel(imds,0.7,'randomized');
net = googlenet;
inputSize = net.Layers(1).InputSize;
lgraph = layerGraph(net); 
numClasses = numel(categories(imdsTrain.Labels))
newLearnableLayer = fullyConnectedLayer(numClasses, ...
    'Name','new_fc', ...
    'WeightLearnRateFactor',10, ...
    'BiasLearnRateFactor',10);
    
lgraph = replaceLayer(lgraph,'loss3-classifier',newLearnableLayer);
newClassLayer = classificationLayer('Name','new_classoutput');
lgraph = replaceLayer(lgraph,'output',newClassLayer);

pixelRange = [-30 30];
imageAugmenter = imageDataAugmenter( ...
    'RandXReflection',true, ...
    'RandXTranslation',pixelRange, ...
    'RandYTranslation',pixelRange);
augimdsTrain = augmentedImageDatastore(inputSize(1:2),imdsTrain, ...
    'DataAugmentation',imageAugmenter);

augimdsValidation = augmentedImageDatastore(inputSize(1:2),imdsValidation);

options = trainingOptions('sgdm', ...
    'MiniBatchSize',10, ...
    'MaxEpochs',6, ...
    'InitialLearnRate',1e-4, ...
    'Shuffle','every-epoch', ...
    'ValidationData',augimdsValidation, ...
    'ValidationFrequency',3, ...
    'Verbose',false, ...
    'Plots','training-progress');
netTransfer = trainNetwork(augimdsTrain,lgraph,options);

Answer 1

[Srivardhan Gadila]

If the classes listed from line "front_or_left = find(imd...." in the above code are the only classes which are 24 in total then the issue might be that the value returned by

numClasses = numel(categories(imdsTrain.Labels))

is probably 28224 and not 24. If that's the case then specify the value 24 directly in the line

newLearnableLayer = fullyConnectedLayer(24, ...

and try running the code.

You can try debugging the issue by checking the layer activation size by using analyzeNetwork function.

If the network is fine then the issue might be w.r.t your data i.e., labels itself.

If you still face the issue then Contact MathWorks Support.

[Abdulaziz Alotaibi]

Thanks a lot!