re training neural network from previous state using trainNetwork

Question

Avi Sulimarski 2021-4-29

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/816775-re-training-neural-network-from-previous-state-using-trainnetwork

评论： Xie Shipley 2023-11-16

采纳的回答： Yomna Genina

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.

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Yomna Genina 2021-5-5

3
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/816775-re-training-neural-network-from-previous-state-using-trainnetwork#answer_692820

在 MATLAB Online 中打开

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!

4 个评论
显示 2更早的评论隐藏 2更早的评论

NASRIN AKTER 2022-6-4

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 2023-11-16

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

请先登录，再进行评论。

re training neural network from previous state using trainNetwork

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

采纳的回答

4 个评论
显示 2更早的评论隐藏 2更早的评论

更多回答（0 个）

另请参阅

类别

标签

产品

版本

Community Treasure Hunt

re training neural network from previous state using trainNetwork

0 个评论 显示 -2更早的评论隐藏 -2更早的评论

采纳的回答

4 个评论 显示 2更早的评论隐藏 2更早的评论

更多回答（0 个）

另请参阅

类别

标签

产品

版本

Community Treasure Hunt

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

4 个评论
显示 2更早的评论隐藏 2更早的评论