Specify Training Options in Custom Training Loop

For most tasks, you can control the training algorithm details using thetrainingOptionsandtrainNetworkfunctions. If thetrainingOptionsfunction does not provide the options you need for your task (for example, a custom learning rate schedule), then you can define your own custom training loop using adlnetworkobject. Adlnetworkobject allows you to train a network specified as a layer graph using automatic differentiation.

To specify the same options as thetrainingOptions, use these examples as a guide:

Training Option	`trainingOptions`Argument	Example
Adam solver	`solverName` `'GradientDecayFactor'` `'SquaredGradientDecayFactor'`	Adaptive Moment Estimation (ADAM)
RMSProp solver	`solverName` `'Epsilon'`	Root Mean Square Propagation (RMSProp)
SGDM solver	`solverName` `'Momentum'`	Stochastic Gradient Descent with Momentum (SGDM)
Learn rate	`“InitialLearnRate'`	Learn Rate
Learn rate schedule	`'LearnRateSchedule'` `“LearnRateDropPeriod”` `'LearnRateDropFactor'`	Piecewise Learn Rate Schedule
Training progress	`“阴谋”`	Plots
Verbose output	`'Verbose'` `'VerboseFrequency'`	Verbose Output
Mini-batch size	`'MiniBatchSize'`	Mini-Batch Size
Number of epochs	`'MaxEpochs'`	Number of Epochs
Validation	`'ValidationData'` `'ValidationPatience'`	Validation
L₂regularization	`'L2Regularization'`	L2 Regularization
Gradient clipping	`'GradientThreshold'` `'GradientThresholdMethod'`	Gradient Clipping
单CPU或GPU培训	`'ExecutionEnvironment'`	Single CPU or GPU Training
Checkpoints	`'CheckpointPath'`	Checkpoints

Solver Options

To specify the solver, use theadamupdate,rmspropupdate, andsgdmupdatefunctions for the update step in your training loop. To implement your own custom solver, update the learnable parameters using thedlupdate函数。

Adaptive Moment Estimation (ADAM)

更新your network parameters using Adam, use theadamupdate函数。Specify the gradient decay and the squared gradient decay factors using the corresponding input arguments.

Root Mean Square Propagation (RMSProp)

更新your network parameters using RMSProp, use thermspropupdate函数。Specify the denominator offset (epsilon) value using the corresponding input argument.

Stochastic Gradient Descent with Momentum (SGDM)

更新your network parameters using SGDM, use thesgdmupdate函数。Specify the momentum using the corresponding input argument.

Learn Rate

To specify the learn rate, use the learn rate input arguments of theadamupdate,rmspropupdate, andsgdmupdatefunctions.

To easily adjust the learn rate or use it for custom learn rate schedules, set the initial learn rate before the custom training loop.

learnRate = 0.01;

Piecewise Learn Rate Schedule

To automatically drop the learn rate during training using a piecewise learn rate schedule, multiply the learn rate by a given drop factor after a specified interval.

To easily specify a piecewise learn rate schedule, create the variableslearnRate,learnRateSchedule,learnRateDropFactor, andlearnRateDropPeriod, wherelearnRateis the initial learn rate,learnRateSchedulecontains either"piecewise"or"none",learnRateDropFactoris a scalar in the range [0, 1] that specifies the factor for dropping the learning rate, andlearnRateDropPeriodis a positive integer that specifies how many epochs between dropping the learn rate.

learnRate = 0.01; learnRateSchedule ="piecewise"learnRateDropPeriod = 10; learnRateDropFactor = 0.1;

Inside the training loop, at the end of each epoch, drop the learn rate when thelearnRateScheduleoption is"piecewise"and the current epoch number is a multiple oflearnRateDropPeriod. Set the new learn rate to the product of the learn rate and the learn rate drop factor.

iflearnRateSchedule = ="piecewise"&& mod(epoch,learnRateDropPeriod) == 0 learnRate = learnRate * learnRateDropFactor;end

Plots

To plot the training loss and accuracy during training, calculate the mini-batch loss and either the accuracy or the root-mean-squared-error (RMSE) in the model loss function and plot them using aTrainingProgressMonitorobject.

To easily specify that the plot should be on or off, set theVisibleproperty of theTrainingProgressMonitorobject. By default,Visibleis set totrue. WhenVisibleis set tofalse,software logs the training metrics and information but does not display the Training Progress window. You can display the Training Progress window after training by changing theVisibleproperty. To also plot validation metrics, use the same optionsvalidationDataandvalidationFrequencydescribed inValidation.

validationData = {XValidation, YValidation}; validationFrequency = 50;

Before training, initialize aTrainingProgressMonitorobject. The monitor automatically tracks the elapsed time since the construction of the object. To use this elapsed time as a proxy for training time, make sure you create theTrainingProgressMonitorobject close to the start of the training loop.

For classification tasks, create a plot to track the loss and accuracy for the training and validation data. Also track the epoch number and the training progress percentage.

monitor = trainingProgressMonitor; monitor.Metrics = ["TrainingAccuracy","ValidationAccuracy","TrainingLoss","ValidationLoss"]; groupSubPlot(monitor,"Accuracy",["TrainingAccuracy","ValidationAccuracy"]); groupSubPlot(monitor,"Loss",["TrainingLoss","ValidationLoss"]); monitor.Info ="Epoch"; monitor.XLabel ="Iteration"; monitor.Progress = 0;

For regression tasks, adjust the code by changing the variable names and labels so that it initializes plots for the training and validation RMSE instead of the training and validation accuracy.

Inside the training loop, at the end of an iteration, use therecordMetricsandupdateInfofunctions to include the appropriate metrics and information for the training loop. For classification tasks, add points corresponding to the mini-batch accuracy and the mini-batch loss. If the current iteration is either 1 or a multiple of the validation frequency option, then also add points for the validation data.

recordMetrics(monitor,iteration,...TrainingLoss=lossTrain,...TrainingAccuracy=accuracyTrain); updateInfo(monitor,Epoch=string(epoch) +" of "+ string(numEpochs));ifiteration == 1 || mod(iteration,validationFrequency) == 0 recordMetrics(monitor,iteration,...ValidationLoss=lossValidation,...ValidationAccuracy=accuracyValidation);endmonitor.Progress = 100*iteration/numIterations;

whereaccuracyTrainandlossTraincorrespond to the mini-batch accuracy and loss calculated in the model loss function. For regression tasks, use the mini-batch RMSE losses instead of the mini-batch accuracies.

You can stop training using theStopbutton in the Training Progress window. When you clickStop,Stopproperty of the monitor changes to1(true). Training stops if your training loop exits when theStopproperty is1.

whilenumEpochs < maxEpochs && ~monitor.Stop% Custom training loop code.end

For more information about plotting and recording metrics during training, seeMonitor Custom Training Loop Progress During Training.

To learn how to compute validation metrics, seeValidation.

Verbose Output

To display the training loss and accuracy during training in a verbose table, calculate the mini-batch loss and either the accuracy (for classification tasks) or the RMSE (for regression tasks) in the model loss function and display them using thedisp函数。

To easily specify that the verbose table should be on or off, create the variablesverboseandverboseFrequency, whereverboseistrueorfalseandverbosefrequencyspecifies how many iterations between printing verbose output. To display validation metrics, use the same optionsvalidationDataandvalidationFrequencydescribed inValidation.

verbose = true verboseFrequency = 50; validationData = {XValidation, YValidation}; validationFrequency = 50;

Before training, display the verbose output table headings and initialize a timer using thetic函数。

disp("|======================================================================================================================|") disp("| Epoch | Iteration | Time Elapsed | Mini-batch | Validation | Mini-batch | Validation | Base Learning |") disp("| | | (hh:mm:ss) | Accuracy | Accuracy | Loss | Loss | Rate |") disp("|======================================================================================================================|") start = tic;

For regression tasks, adjust the code so that it displays the training and validation RMSE instead of the training and validation accuracy.

Inside the training loop, at the end of an iteration, print the verbose output when theverboseoption istrueand it is either the first iteration or the iteration number is a multiple ofverboseFrequency.

ifverbose && (iteration == 1 || mod(iteration,verboseFrequency) == 0 D = duration(0,0,toc(start),'Format','hh:mm:ss');ifisempty(validationData) || mod(iteration,validationFrequency) ~= 0 accuracyValidation =""; lossValidation ="";enddisp("| "+...pad(epoch,7,'left') +" | "+...pad(iteration,11,'left') +" | "+...pad(D,14,'left') +" | "+...pad(accuracyTrain,12,'left') +" | "+...pad(accuracyValidation,12,'left') +" | "+...pad(lossTrain,12,'left') +" | "+...pad(lossValidation,12,'left') +" | "+...pad(learnRate,15,'left') +" |")end

For regression tasks, adjust the code so that it displays the training and validation RMSE instead of the training and validation accuracy.

When training is finished, print the last border of the verbose table.

disp("|======================================================================================================================|")

To learn how to compute validation metrics, seeValidation.

Mini-Batch Size

Setting the mini-batch size depends on the format of data or type of datastore used.

To easily specify the mini-batch size, create a variableminiBatchSize.

miniBatchSize = 128;

For data in an image datastore, before training, set theReadSizeproperty of the datastore to the mini-batch size.

imds.ReadSize = miniBatchSize;

For data in an augmented image datastore, before training, set theMiniBatchSizeproperty of the datastore to the mini-batch size.

augimds.MiniBatchSize = miniBatchSize;

For in-memory data, during training at the start of each iteration, read the observations directly from the array.

idx = ((iteration - 1)*miniBatchSize + 1):(iteration*miniBatchSize); X = XTrain(:,:,:,idx);

Number of Epochs

Specify the maximum number of epochs for training in the outer loop of the training loop.

To easily specify the maximum number of epochs, create the variablemaxEpochsthat contains the maximum number of epochs.

maxEpochs = 30;

In the outer loop of the training loop, specify to loop over the range 1, 2, …,maxEpochs.

epoch = 0;whilenumEpochs < maxEpochs epoch = epoch + 1;...end

Validation

To validate your network during training, set aside a held-out validation set and evaluate how well the network performs on that data.

To easily specify validation options, create the variablesvalidationDataandvalidationFrequency, wherevalidationDatacontains the validation data or is empty andvalidationFrequencyspecifies how many iterations between validating the network.

validationData = {XValidation,TValidation}; validationFrequency = 50;

During the training loop, after updating the network parameters, test how well the network performs on the held-out validation set using thepredict函数。Validate the network only when validation data is specified and it is either the first iteration or the current iteration is a multiple of thevalidationFrequencyoption.

ifiteration == 1 || mod(iteration,validationFrequency) == 0 YValidation = predict(net,XValidation); lossValidation = crossentropy(YValidation,TValidation); [~,idx] = max(YValidation); labelsPredValidation = classNames(idx); accuracyValidation = mean(labelsPredValidation == labelsValidation);end

Here,TValidationis a one-hot encoded array of labels overclassNames. To calculate the accuracy, convertTValidationto an array of labels.

For regression tasks, adjust the code so that it calculates the validation RMSE instead of the validation accuracy.

For an example showing how to calculate and plot validation metrics during training, seeMonitor Custom Training Loop Progress During Training.

Early Stopping

To stop training early when the loss on the held-out validation stops decreasing, use a flag to break out of the training loops.

To easily specify the validation patience (the number of times that the validation loss can be larger than or equal to the previously smallest loss before network training stops), create the variablevalidationPatience.

validationPatience = 5;

Before training, initialize a variablesearlyStopandvalidationLosses, whereearlyStopis a flag to stop training early andvalidationLossescontains the losses to compare. Initialize the early stopping flag withfalseand array of validation losses withinf.

earlyStop = false;ifisfinite(validationPatience) validationLosses = inf(1,validationPatience);end

Inside the training loop, in the loop over mini-batches, add theearlyStop旗帜the loop condition.

whilehasdata(ds) && ~earlyStop...end

During the validation step, append the new validation loss to the arrayvalidationLosses. If the first element of the array is the smallest, then set theearlyStop旗帜true. Otherwise, remove the first element.

ifisfinite(validationPatience) validationLosses = [validationLosses validationLoss];ifmin(validationLosses) == validationLosses(1) earlyStop = true;elsevalidationLosses(1) = [];endend

L₂Regularization

To apply L₂regularization to the weights, use thedlupdate函数。

To easily specify the L₂regularization factor, create the variablel2Regularizationthat contains the L₂regularization factor.

l2Regularization = 0.0001;

During training, after computing the model loss and gradients, for each of the weight parameters, add the product of the L₂regularization factor and the weights to the computed gradients using thedlupdate函数。更新only the weight parameters, extract the parameters with name"Weights".

idx = net.Learnables.Parameter =="Weights"; gradients(idx,:) = dlupdate(@(g,w) g + l2Regularization*w, gradients(idx,:), net.Learnables(idx,:));

After adding the L₂regularization parameter to the gradients, update the network parameters.

Gradient Clipping

To clip the gradients, use thedlupdate函数。

To easily specify gradient clipping options, create the variablesgradientThresholdMethodandgradientThreshold, wheregradientThresholdMethodcontains"global-l2norm","l2norm", or"absolute-value", andgradientThresholdis a positive scalar containing the threshold orinf.

gradientThresholdMethod ="global-l2norm"; gradientThreshold = 2;

Create functions namedthresholdGlobalL2Norm,thresholdL2Norm, andthresholdAbsoluteValuethat apply the"global-l2norm","l2norm", and"absolute-value"threshold methods, respectively.

For the"global-l2norm"option, the function operates on all gradients of the model.

functiongradients = thresholdGlobalL2Norm(gradients,gradientThreshold) globalL2Norm = 0;fori = 1:numel(gradients) globalL2Norm = globalL2Norm + sum(gradients{i}(:).^2);endglobalL2Norm = sqrt(globalL2Norm);ifglobalL2Norm > gradientThreshold normScale = gradientThreshold / globalL2Norm;fori = 1:numel(gradients) gradients{i} = gradients{i} * normScale;endendend

For the"l2norm"and"absolute-value"options, the functions operate on each gradient independently.

functiongradients = thresholdL2Norm(gradients,gradientThreshold) gradientNorm = sqrt(sum(gradients(:).^2));ifgradientNorm > gradientThreshold gradients = gradients * (gradientThreshold / gradientNorm);endend

functiongradients = thresholdAbsoluteValue(gradients,gradientThreshold) gradients(gradients > gradientThreshold) = gradientThreshold; gradients(gradients < -gradientThreshold) = -gradientThreshold;end

During training, after computing the model loss and gradients, apply the appropriate gradient clipping method to the gradients using thedlupdate函数。因为"global-l2norm"option requires all the gradient values, apply thethresholdGlobalL2Normfunction directly to the gradients. For the"l2norm"and"absolute-value"options, update the gradients independently using thedlupdate函数。

switchgradientThresholdMethodcase"global-l2norm"gradients = thresholdGlobalL2Norm(gradients, gradientThreshold);case"l2norm"gradients = dlupdate(@(g) thresholdL2Norm(g, gradientThreshold),gradients);case"absolute-value"gradients = dlupdate(@(g) thresholdAbsoluteValue(g, gradientThreshold),gradients);end

After applying the gradient threshold operation, update the network parameters.

Single CPU or GPU Training

The software, by default, performs calculations using only the CPU. To train on a single GPU, convert the data togpuArrayobjects.Using a GPU requires a Parallel Computing Toolbox™ license and a supported GPU device. For information about supported devices, seeGPU Computing Requirements(Parallel Computing Toolbox).

To easily specify the execution environment, create the variableexecutionEnvironmentthat contains either"cpu","gpu", or"auto".

executionEnvironment ="auto"

During training, after reading a mini-batch, check the execution environment option and convert the data to agpuArrayif necessary. ThecanUseGPUfunction checks for useable GPUs.

if(executionEnvironment =="auto"&& canUseGPU) || executionEnvironment =="gpu"X = gpuArray(X);end

Checkpoints

To save checkpoint networks during training save the network using thesave函数。

To easily specify whether checkpoints should be switched on, create the variablecheckpointPathcontains the folder for the checkpoint networks or is empty.

checkpointPath = fullfile(tempdir,"checkpoints");

If the checkpoint folder does not exist, then before training, create the checkpoint folder.

if~exist(checkpointPath,"dir") mkdir(checkpointPath)end

During training, at the end of an epoch, save the network in a MAT file. Specify a file name containing the current iteration number, date, and time.

if~isempty(checkpointPath) D = string(datetime("now",Format="yyyy_MM_dd__HH_mm_ss")); filename ="net_checkpoint__"+ iteration +"__"+ D +".mat"; save(filename,"net")end

wherenetis thedlnetworkobject to be saved.