How can I modify the first layer of a pre-trained DNN

I am using a DNN(resnet-50) for feature extraction. However, the visual features of a category are very small and the resizing of the features delete the data. CNN and pool layer do not get affected by dimmension size change.

When i try to change the input size i get the error:

net.Layers(1).InputSize=[280,280,3]

You cannot set the read-only property 'InputSize' of ImageInputLayer.

How can i modify this layer?

Answer 1

[Wilmer Ariza]

I found the solution,

The network has to be move to layer graph, replace layer then can be applied.

I am attaching the code for feature extraction example from amtlab with modified input size

unzip('MerchData.zip');
imds = imageDatastore('MerchData','IncludeSubfolders',true,'LabelSource','foldernames');
[imdsTrain,imdsTest] = splitEachLabel(imds,0.7,'randomized');


numTrainImages = numel(imdsTrain.Labels);
idx = randperm(numTrainImages,16);
figure
for i = 1:16
    subplot(4,4,i)
    I = readimage(imdsTrain,idx(i));
    imshow(I)
end



net = resnet18;
analyzeNetwork(net)

inputSize = net.Layers(1).InputSize;
% create layer that is going to replace input layer

% the size of your image must be bigger or equal to the input size.
layer=imageInputLayer([28 28 3],'Name','input','Normalization','zscore');
% move data for normalization and mean
layer.Mean=lgraph.Layers(1,1).Mean;
layer.StandardDeviation=lgraph.Layers(1, 1).StandardDeviation;
%% Move network to layer graph
lgraph = layerGraph(net);
newlgraph = replaceLayer(lgraph,'data',layer);
% Assemble DNN to work again with it
net = assembleNetwork(newlgraph);
analyzeNetwork(net)

augimdsTrain = augmentedImageDatastore(inputSize(1:2),imdsTrain);
augimdsTest = augmentedImageDatastore(inputSize(1:2),imdsTest);

layer = 'pool5';
featuresTrain = activations(net,augimdsTrain,layer,'OutputAs','channels');

featuresTrain = squeeze(featuresTrain);

featuresTrain=featuresTrain';
featuresTest = activations(net,augimdsTest,layer,'OutputAs','channels');
featuresTest = squeeze(featuresTest);

featuresTest=featuresTest';
whos featuresTrain

analyzeNetwork(net)
YTrain = imdsTrain.Labels;
YTest = imdsTest.Labels;


classifier = fitcecoc(featuresTrain,YTrain);


YPred = predict(classifier,featuresTest);

idx = [1 5 10 15];
figure
for i = 1:numel(idx)
    subplot(2,2,i)
    I = readimage(imdsTest,idx(i));
    label = YPred(idx(i));
    imshow(I)
    title(char(label))
end


accuracy = mean(YPred == YTest)