Extracting feature vectors as input to train other network

After Extracting feature vectors from pre-trained models, I would like to use those produced features to train a deep neural network with a number of fully connected layers. What can i do?

Thank you for your answer.

Answer 1

[Animesh Gupta]

Hello,

It is my understanding that you want to reuse the feature vector from pre-trained model and then append it with fully connected layers to train a custom deep learning model.

You may refer the following script that demonstrates a similar procedure.

In this demonstrattion, we are using pretrained GoolgeLeNet neural network and replacing fully connected layer and output layer.

net = googlenet; % loading pretrained GoogleLeNet neural network 
lgraph = layerGraph(net);  % extracting the layer graph of the model

newLearnableLayer = fullyConnectedLayer(5, ...
    'Name','new_fc', ...
    'WeightLearnRateFactor',10, ...
    'BiasLearnRateFactor',10); % creating a new custom layer for our model

lgraph = replaceLayer(lgraph,'loss3-classifier',newLearnableLayer); % replace the existing 'loss3-classifier' with our newLearnableLayer

newClassLayer = classificationLayer('Name','new_classoutput'); % creating a new classification layer of name as new_classoutput
lgraph = replaceLayer(lgraph,'output',newClassLayer); % replacing output layer of original GoogleLeNet with our custom output layer

deepNetworkDesigner(lgraph) % visualizing the change

It can be observed that "loss3-classifier" and "output" layers are replaced with our new custom layers.

In a similar fashion, new layers can also be added in the network using addLayers and connectLayers method of layerGraoh object.

You can refer the following documentation for more information -

Layer Graph - https://www.mathworks.com/help/deeplearning/ref/nnet.cnn.layergraph.html

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

I hope it helps.