Hello,
My understanding is that you are trying to use the “insightFace” File Exchange program and would like to understand the classification layer and details on how to perform the classification.
The code below is part of the source code for “insightFace” and models the layer structure of the neural network. The ‘embeddingL’ layer is replacing the Classfication layer of the original ‘mobilenetv2’ backbone.
%% network arc
backbone = mobilenetv2();
lg = layerGraph(backbone);
lg = removeLayers(lg,{'input_1',...
'Logits','Logits_softmax',...
'ClassificationLayer_Logits'});
inputLayer = imageInputLayer([inputSize,3],'name','input','Normalization','none');
embeddingL = fullyConnectedLayer(embedding_size,'name','embedding',...
'Bias',zeros(embedding_size,1,'single'),'BiasLearnRateFactor',0);% embedding
lg = addLayers(lg,inputLayer);
lg = addLayers(lg,embeddingL);
lg = connectLayers(lg,'input','Conv1');
lg = connectLayers(lg,'global_average_pooling2d_1','embedding');
A ‘dlnetwork’ object is then used to train this model. The following documentation link contains the end-to-end workflow for training and testing a model using ‘dlnetwork’ object.
As an alternative you can consider exploring the File Exchange program ‘Implementation-ArcFace-in-MATLAB’ which contains a working example for the Oxford-IIIT Pet Datset. Please refer to the link below.
I hope this helps.
