Hi Tan,
You are extracting layers from ‘GoogleNet’ in the following code and then using the extracted layer to form a ‘layerGraph’ and then further add/replace other layers like fully connection layer and classification layer.
%Load pre-trained network
net=googlenet;
analyzeNetwork(net)
lys = net.Layers;
lys(end-3:end)
numClasses = numel(categories(imdsTrain.Labels));
lgraph = layerGraph(lys);
However, ‘Googlenet’ has multiple parallel connections from a single layer throughout the network. When you are create layerGraph from net.Layers, connection information in ‘net.Connections’ is not being used. ‘layerGraph’ can directly be created from network by ‘lgraph = layerGraph(net)’ so that connection information is also there. You may use the following code:
net=googlenet;
analyzeNetwork(net)
numClasses = 5;
lgraph = layerGraph(net);
% check the layer connected to fully connected layer]
inputSizeFc = lgraph.Layers(142).InputSize;
%Replace the classification layers for new task
newFCLayer = fullyConnectedLayer(numClasses,'Name','new_fc','WeightLearnRateFactor',10,'BiasLearnRateFactor',10, ...
'Bias',ones(numClasses, 1), 'Weights',ones(numClasses, inputSizeFc));
lgraph = replaceLayer(lgraph,'loss3-classifier',newFCLayer);
newClassLayer = classificationLayer('Name','new_classoutput');
lgraph = replaceLayer(lgraph,'output',newClassLayer);
analyzeNetwork(lgraph);
