trainnet function - randomness and repeatability

Hello all,

I'm checking the trainnet function. I'm running same script multiple times and training outputs are slightly different. Even though I remove all points which brings any randomness in (I'm aware about) - like random label splitting, batches schuffling, etc.

So may I ask you, does anybody know what causes that outcomes are slightly different each time? Please find below basic steps of my script.

training_imds = imageDatastore(Training_data_folder,"IncludeSubfolders",true,"Labelsource","foldernames");

%-----------------------
% Training data split
% only part of training dataset can be used for the training
% training_imdsVal is not used
[training_imds_Train,training_imdsVal] = splitEachLabel(training_imds,0.3);
training_imds = training_imds_Train;
%-----------------------

% Training process - train - val data split
[training_imds_Train,training_imds_Val] = splitEachLabel(training_imds,0.9);

training_imds_Train_au = augmentedImageDatastore([imHeight imWidth],training_imds_Train);
training_imds_Val_au   = augmentedImageDatastore([imHeight imWidth],training_imds_Val);

layers = [
    imageInputLayer([imHeight imWidth 3]) % image size and RGB (=3)
    convolution2dLayer(20,20)
    reluLayer()
    maxPooling2dLayer(3)
    fullyConnectedLayer(2)
    softmaxLayer()
];

options = trainingOptions("sgdm", ...
    Metrics="accuracy", ...
    InitialLearnRate=0.000001, ...
    ValidationData=training_imds_Val_au,...
    MiniBatchSize=128,... 
    ValidationFrequency=25,...
    ValidationPatience=5,...
    MaxEpochs = 1,...
    LearnRateSchedule = 'piecewise',... 
    LearnRateDropPeriod = 5,...
    ExecutionEnvironment='cpu');

trained_net = trainnet(training_imds_Train_au,layers,"crossentropy",options);

When I run this more times, different outcomes are received, e.g.:

    Iteration    Epoch    TimeElapsed    LearnRate    TrainingLoss    ValidationLoss    TrainingAccuracy    ValidationAccuracy
    _________    _____    ___________    _________    ____________    ______________    ________________    __________________
            0        0       00:00:10        1e-06                            1.4484                                    81.864
            1        1       00:00:10        1e-06          5.8723                                60.156                      
           25        1       00:01:31        1e-06          1.3652           0.40944              91.406                97.229
           50        1       00:03:17        1e-06         0.12455         1.051e-09              99.219                   100
           55        1       00:03:43        1e-06         0.16857                 0              98.438                   100
Training stopped: Max epochs completed

    Iteration    Epoch    TimeElapsed    LearnRate    TrainingLoss    ValidationLoss    TrainingAccuracy    ValidationAccuracy
    _________    _____    ___________    _________    ____________    ______________    ________________    __________________
            0        0       00:00:09        1e-06                            7.5433                                    51.637
            1        1       00:00:10        1e-06          8.6849                                41.406                      
           25        1       00:01:27        1e-06          0.9964           0.14055               93.75                99.118
           50        1       00:03:03        1e-06         0.62306          0.022228              96.094                99.748
           55        1       00:03:23        1e-06         0.12455          0.009824              99.219                99.874
Training stopped: Max epochs completed

Answer 1

[Steven Lord]

What happens when you reset the state or seed of the random number generator before each attempt to train the network? Let's choose an arbitrary seed value and generate some numbers.

rng(42)
x1 = rand(1, 5);

If we reset the seed to the same value, the generator starts in the same place and generates the same numbers.

rng(42)
x2 = rand(1, 5);
isequal(x1, x2) % Same values, down to the last bit

ans = logical
   1

But generating new values doesn't generate the same values as the freshly-reset generator.

x3 = rand(1, 5);
isequal(x1, x3) % No, x3 contains different values

ans = logical
   0

You may have removed the randomness from your code, but I believe the network may be initialized with random starting values for the training internally.

Answer 2

[Matt J]

In addition to random initialization of the Learnables, as mentioned by @Steven Lord, you are using the default Shuffle setting, which performs a random reordering of the training inputs once at the beginning of the training process.