How to plot training loss for Covolutional Neural Nets?

To plot training loss for a Convolutional Neural Network (CNN) in MATLAB, set 'Plots', 'training-progress' inside trainingOptions() to view live real-time loss and accuracy curves. To generate custom publication-ready plots after training, capture the training output struct using [net, info] = trainNetwork(...) and plot info.TrainingLoss directly.

Two Primary Plotting Approaches

1. Live Interactive Monitor: Adding 'Plots', 'training-progress' launches MATLAB's dedicated Deep Learning Monitor window. It plots mini-batch loss, smoothed loss, mini-batch accuracy, validation metrics, and elapsed training time on the fly.
2. Post-Training Custom Figure: The info structure returned by trainNetwork stores numerical arrays for TrainingLoss, TrainingAccuracy, ValidationLoss, and ValidationAccuracy. This allows full control over fonts, axis limits, line weights, and multi-axis formatting for IEEE or Elsevier papers.

Executable MATLAB Code: Live Monitoring & Custom Loss Curves

% =========================================================================
% Plotting CNN Training Loss and Accuracy in MATLAB
% =========================================================================

% 1. Load sample digit image data
[XTrain, YTrain] = digitTrain4DArrayData;
[XValidation, YValidation] = digitTest4DArrayData;

% 2. Define a simple Convolutional Neural Network architecture
layers = [
    imageInputLayer([28 28 1], 'Name', 'input')
    
    convolution2dLayer(3, 16, 'Padding', 'same', 'Name', 'conv1')
    batchNormalizationLayer('Name', 'bn1')
    reluLayer('Name', 'relu1')
    maxPooling2dLayer(2, 'Stride', 2, 'Name', 'pool1')
    
    convolution2dLayer(3, 32, 'Padding', 'same', 'Name', 'conv2')
    batchNormalizationLayer('Name', 'bn2')
    reluLayer('Name', 'relu2')
    
    fullyConnectedLayer(10, 'Name', 'fc')
    softmaxLayer('Name', 'softmax')
    classificationLayer('Name', 'output')
];

% 3. Configure training options with both Live Plotting and Validation Tracking
options = trainingOptions('sgdm', ...
    'InitialLearnRate', 0.01, ...
    'MaxEpochs', 6, ...
    'MiniBatchSize', 128, ...
    'Shuffle', 'every-epoch', ...
    'ValidationData', {XValidation, YValidation}, ...
    'ValidationFrequency', 20, ...
    'Plots', 'training-progress', ... % Live real-time plot window
    'Verbose', false);

% 4. Train the network and capture the 'info' structure
[net, info] = trainNetwork(XTrain, YTrain, layers, options);

% =========================================================================
% 5. Create a Custom Publication-Quality Loss & Accuracy Plot
% =========================================================================

figure('Color', 'w', 'Position', [100, 100, 800, 450]);

% Left Y-Axis: Training and Validation Loss
yyaxis left
plot(info.TrainingLoss, 'LineWidth', 1.2, 'Color', [0.85, 0.32, 0.1]);
hold on;
valIterations = 1:options.ValidationFrequency:length(info.TrainingLoss);
valLoss = info.ValidationLoss(~isnan(info.ValidationLoss));
valIterIndices = linspace(1, length(info.TrainingLoss), length(valLoss));

plot(valIterIndices, valLoss, 'o--', 'LineWidth', 1.5, ...
    'MarkerSize', 5, 'Color', [0.63, 0.08, 0.18]);
ylabel('Cross-Entropy Loss', 'FontWeight', 'bold');
ylim([0, max(info.TrainingLoss) * 1.1]);

% Right Y-Axis: Training and Validation Accuracy
yyaxis right
smoothedTrainAcc = movmean(info.TrainingAccuracy, 5); % 5-iteration smoothing
plot(smoothedTrainAcc, 'LineWidth', 1.2, 'Color', [0.0, 0.45, 0.74]);
valAcc = info.ValidationAccuracy(~isnan(info.ValidationAccuracy));
plot(valIterIndices, valAcc, 's--', 'LineWidth', 1.5, ...
    'MarkerSize', 5, 'Color', [0.47, 0.67, 0.19]);
ylabel('Accuracy (%)', 'FontWeight', 'bold');
ylim([0, 105]);

grid on;
xlabel('Iteration', 'FontWeight', 'bold');
title('CNN Training Loss and Classification Accuracy Convergence');
legend({'Training Loss', 'Validation Loss', 'Training Acc (Smoothed)', 'Validation Acc'}, ...
    'Location', 'east');

% Optional: Export vector graphic for papers
% exportgraphics(gcf, 'cnn_loss_curve.eps', 'Resolution', 300);

Key Details for Technical Optimization

• Smoothing High-Frequency Jitter: Mini-batch loss frequently fluctuates between individual batches. Use movmean(info.TrainingLoss, 10) to calculate an exponential or moving average baseline that reveals the true underlying learning trajectory.
• Handling NaN Values in Validation Vectors: Because validation is evaluated periodically rather than every iteration, the info.ValidationLoss array contains NaN values on non-validation steps. Always filter them using ~isnan(info.ValidationLoss) before plotting lines.
• Custom Training Loops (dlnetwork): If training with custom loss functions via dlgradient and dlfeval instead of trainNetwork, use MATLAB's animatedline and addpoints inside the epoch loop to stream loss curves directly to an active figure.