How to retrain segnet?

Hello there, i am trying to do an incremental training of segnet available in Matlab (train first on a small database and then use resultant net to train on larger data). This i am not able to do with segnet as resultant net is not trainable network format. Here is the code i used,

lgraph = segnetLayers(imageSize, numClasses, 4); net = trainNetwork(trainingData, lgraph, options); save seg_net net; % net = retrain(trainingData, net, options); %what to write in retrain?!

Answer 1

[Purvaja]

Hi @vipin xavior,

I understand that you want to achieve incremental training of an already trained “segnetLayers” model and train it on a different dataset.

We can approach this by first saving, loading and then converting it to layers since “net.Layers” object does not retain the connections between layers. So, we will construct lost connections using “lgraph” from the saved network and then use it for incremental training.

You can follow the given procedures:

• Create the SegNet layers

lgraph = segnetLayers(imageSize,numClasses,2); 

• Train the network

net = trainNetwork(ds, lgraph, options); 

• Save the trained network

save('segnet.mat', 'net'); 

• Load the saved network in another file

load('segnet.mat', 'net'); 

• Reconstruct the layer graph from the loaded network

lgraph = layerGraph(net); 

• Incrementally train the network

net = trainNetwork(ds, lgraph, incrementalOptions); 

For more clarification, the following documentation links would be helpful:

1. This link will give you information about “segnetLayers” function, https://www.mathworks.com/help/vision/ref/segnetlayers.html
2. This link describes about “layerGraph” that will help in building lost connections, https://www.mathworks.com/help/deeplearning/ref/nnet.cnn.layergraph.html

Since MATLAB does not recommend using “LayerGraph” and “segnetLayers” as they would be deprecated in the future, you can explore “dlnetwork”.

https://www.mathworks.com/help/deeplearning/ref/dlnetwork.html

Hope this helps you!