Automatically stop training of neural network if certain accuracy in training is reached

I would like to do a sensitivity analysis to get the best parameters of my simulated training data. To make it automatic i would like the training to stop automatically when the network reaches an accuracy of x% on the training dataset. Is there an option or parameter for this?

Currently my code looks like this:

layers = [...
    imageInputLayer([100,100,1])
    convolution2dLayer(3,4,'Stride',1,'Padding',1)
    batchNormalizationLayer
    reluLayer
    maxPooling2dLayer(2,'Stride',2,'Padding',0)
    convolution2dLayer(3,8,'Stride',1,'Padding',1)
    batchNormalizationLayer
    reluLayer
    maxPooling2dLayer(2,'Stride',2,'Padding',0)
    fullyConnectedLayer(32)
    fullyConnectedLayer(32)       
    fullyConnectedLayer(2)
    softmaxLayer
    classificationLayer];

options = trainingOptions('sgdm','InitialLearnRate',0.001,'MaxEpochs',20,'Plots','none','MiniBatchSize',500);

net = trainNetwork(trainingHeatmapComplete4d,trainingOutputsComplete,layers,options);

Answer 1

[Milan Bansal]

Hi,

In my understanding, you want to interrupt the training of Neural Network automatically when a certain accuracy is reached.

It is possible to do so by defining a callback function and setting it as "OutputFcn" in "options" argument of the "trainNetwork" function. Add the following function at the bottom of the script.

function stop = accuracyCallback(info)
    stop = false;  % Initialize the stop flag as false
    if info.TrainingAccuracy > 90 % Desired accuracy
        stop = true;  % Set the stop flag to true if desired accuracy is reached
    end
end

This function will act as callback function to stop the training at 90% accuracy.

Set the 'OutputFcn' option to the callback function defined above in the "options" argument as shown below.

options = trainingOptions('sgdm','InitialLearnRate',0.001, ...
    'MaxEpochs',20,'Plots','none','MiniBatchSize',500, ...
    'OutputFcn', @accuracyCallback);

Refer to the documentation link to know more about "trainNetwork" function.

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