re training neural network from previous state using trainNetwork

Hi

I am training a deep neural network , using the following matlab function:

net = trainNetwork(XTrain,YTrain,layers,options);

could I use the trainNetwork command to retrain the network (not from scratch), using the last network state from previous training?

I am sharing some of the code:

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

layers = [

imageInputLayer([1 Nin 5],"Name","imageinput")

convolution2dLayer([1 3],32,"Name","conv","Padding","same")%

%batchNormalizationLayer('Name','batchDown')

tanhLayer("Name","tanh1")

convolution2dLayer([1 3],32,"Name","conv","Padding","same","DilationFactor",[1 3])%

%batchNormalizationLayer('Name','batchDown1')

tanhLayer("Name","tanh2")

convolution2dLayer([1 3],2,"Name","conv","Padding","same","DilationFactor",[1 9])%

regressionLayer];

options = trainingOptions('adam', ...

'InitialLearnRate',0.001, ...

'MaxEpochs',50000, ...

'ExecutionEnvironment','parallel',...

'Verbose',false, ...

'Plots','training-progress');

Net1 = trainNetwork(XTrain1,YTrain1,layers,options);

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

now I would like to train again getting Net2 using new data, and starting the training from Net1 stage.

for example:

Net2 = trainNetwork(XTrain2,YTrain2,layers,options);

however its not clear to me how to start the train process for Net2 using Net1 stage.

Answer 1

[Yomna Genina]

You can do the following if Net1 is a SeriesNetwork:

Net2 = trainNetwork(XTrain2, YTrain2, Net1.Layers, options);

or if Net1 is a DAGNetwork:

Net2 = trainNetwork(XTrain2, YTrain2, layerGraph(Net1), options);

This will train using Net1 as the initial network.

If you would also like to prevent weights in certain layers from changing, you could use freezeWeights (see how to access it below). This function could be used to set the learning rates in those layers to zero. During training, trainNetwork does not update the parameters of the "frozen" layers.

edit(fullfile(matlabroot,'examples','nnet','main','freezeWeights.m'))

You might find the following documentation pages useful:

https://www.mathworks.com/help/deeplearning/ug/resume-training-from-a-checkpoint-network.html

https://www.mathworks.com/help/deeplearning/ug/train-deep-learning-network-to-classify-new-images.html#TransferLearningUsingGoogLeNetExample-4

Hope this helps!

[Avi Sulimarski]

Hi Yomna ,I found your answer very helpful. thanks for the clarification .

[ali ganjbakhsh]

thank you very much, it works.

[NASRIN AKTER]

Hi Yomna

Is it necessary to save the checkpoint networks as shown in the 1st link?

Also can you suggest any example of incremental learning for image classification using deep learning?

Thanks

[Xie Shipley]

If using "population" in `trainingOption`,will the mean and variance of the Net2(it's mean and variance is same as Net1 when stop training Net1) changing while training ? @Yomna Genina