Matlab Train Network

matlab-train-network

Train, evaluate, and export neural networks to Simulink in MATLAB using the recommended dlnetwork-based API (trainnet, dlnetwork, minibatchpredict, scores2label, testnet, imagePretrainedNetwork) or, for tabular data, the Statistics and Machine Learning Toolbox functions fitcnet and fitrnet.

When to Use

Activate this skill when a user asks to:

• Train any neural network (classifier, regression, multi-output, LSTM, CNN, etc.)
• Fine-tune or use a pretrained model for transfer learning
• Evaluate a trained network on test data
• Run inference / predict with a trained network
• Export a trained network to Simulink
• Migrate existing legacy (patternnet, fitnet, narxnet, gensim) or discouraged (trainNetwork, DAGNetwork, classify) code to recommended APIs
• Create a "pattern recognition network", "function fitting network", "NARX network", or any task historically associated with the Neural Network Toolbox shallow nets API
• Speed up or optimize any deep learning code (even without mentioning dlaccelerate by name)
• Make existing deep learning code faster using dlaccelerate
• Diagnose and fix dlaccelerate issues (low HitRate, retracing, code is slower after using dlaccelerate)
• Accelerate custom training (code that uses dlfeval/dlgradient)
• Accelerate a function that supports dlarray input and is long running
• Accelerate a custom loss function passed to trainnet (R2026a+)

When NOT to Use

• Importing/exporting models (importNetworkFromPyTorch, exportONNXNetwork)
• Data loading and preprocessing (imageDatastore, transforms, augmentation)
• Network architecture design decisions (choosing CNN vs LSTM vs transformer)
• Reinforcement learning workflows (use Reinforcement Learning Toolbox)
• Object detection (use specialized detector training functions in Computer Vision Toolbox)

Decision: fitrnet/fitcnet or trainnet

Apply this check before starting any training workflow below.

Pass tables directly to trainnet, fitcnet, and fitrnet. If inputs have categorical columns, pass them directly — they are encoded automatically (fitcnet/fitrnet always; trainnet/minibatchpredict/testnet from R2025a).

% Classification
mdl = fitcnet(tbl,responseName,LayerSizes=20);
[labels,score] = predict(mdl,tblTest);
L = loss(mdl,tblTest);

% Regression
mdl = fitrnet(tbl,responseName,LayerSizes=[20 20]);
Y = predict(mdl,tblTest);
L = loss(mdl,tblTest);

% Tabular data with trainnet (when fitcnet/fitrnet can't be used)
net = trainnet(tbl,net,"crossentropy",options);
accuracy = testnet(net,tblTest,"accuracy");
scores = minibatchpredict(net,tblPredictors);

• From R2024b, fitrnet supports multi-response variables.
• From R2025a, for custom architectures beyond LayerSizes, Activations, LayerWeightsInitializer, and LayerBiasesInitializer, pass a dlnetwork via the Network name-value argument.

Conventions

Training with trainnet + dlnetwork

Data formats

trainnet expects data in specific orientations by default:

If your data has a different layout, use InputDataFormats and/or TargetDataFormats in trainingOptions instead of transposing the data manually. The format string describes your data's current layout — one letter per dimension, not the desired layout. MATLAB handles the remapping internally. For cell arrays, add "B" (batch) to the format string — e.g., InputDataFormats="CTB" for cells of C×T matrices. Do not specify these options when data already matches the input layer's default.

What trainnet supports

Use trainnet and dlnetwork for all Deep Learning Toolbox training. This includes:

• Standard classification and regression
• Transfer learning
• Multi-input or multi-output networks
• Custom loss functions (pass a function handle to trainnet)
• Custom loss function backward passes via DifferentiableFunction
• Custom metrics (string, function handle, or deep.Metric subclass)
• Custom stopping criteria via OutputFcn in trainingOptions
• Custom layers

Custom training loops (dlfeval/dlgradient/update functions) are appropriate when the workflow requires customizations impossible via trainingOptions or the specific workflow — multi-model adversarial training, alternating updates, or custom weight update rules. Note that trainingOptions supports L-BFGS (R2023b+) and Levenberg-Marquardt "lm" (R2024b+).

When a user has a working custom training loop and asks to speed it up, apply dlaccelerate directly. Mention that their workflow may also be expressible with trainnet (which handles acceleration internally), but do not push the conversion — focus on accelerating the code they have.

NEVER use these legacy or discouraged APIs

If the user has existing code using these APIs, migrate it to the recommended replacement and briefly explain which APIs were replaced and what the modern equivalents are. If the user asks for a legacy or discouraged API by name, acknowledge their request and explain that the function has been replaced with a recommended alternative before providing the solution.

See references/legacy-api-redirects.md for before/after code examples.

Inference — use minibatchpredict (or predict)

If you see a for-loop calling predict or forward on batches for inference, replace the entire loop with minibatchpredict. It handles batching, GPU transfer, dlarray conversion, and acceleration automatically. Output format (numeric array, table, or cell array) depends on the input type and network.

Exception: If the loop includes custom pre- or postprocessing around the predict call that cannot be separated from it, minibatchpredict cannot replicate the full pipeline. In that case, wrap the entire custom function with dlaccelerate instead (see references/dlaccelerate-workflow.md). Do not split the function into a minibatchpredict call plus separate accelerated pre/postprocessing — a single dlaccelerate boundary around the full function produces one unified trace.

• For classification: use minibatchpredict (or predict) + scores2label.
• For regression or when you need raw scores: use minibatchpredict or predict.
• predict is for single-batch/small-batch use and accepts plain numeric arrays directly — do not wrap inputs in dlarray or call extractdata on outputs.

Evaluation — use testnet

• Use testnet to calculate post-training metrics on a test dataset instead of doing it manually.
• For single-output networks, use string metrics: "accuracy", "rmse".
• trainnet and testnet accept targets as a separate argument only for in-memory data (testnet(net,XTest,TTest,"accuracy")). When passing a datastore, targets must already be embedded in it (e.g., labeled imageDatastore or combined datastore with targets in a second column).
• For multi-output networks or advanced metric customization, see references/metrics-guidance.md.

Transfer learning — use imagePretrainedNetwork

net = imagePretrainedNetwork("squeezenet",NumClasses=5);

options = trainingOptions("adam", ...
    MaxEpochs=10, ...
    MiniBatchSize=16, ...
    InitialLearnRate=1e-4, ...
    ValidationData=imdsVal, ...
    Metrics="accuracy", ...
    Plots="training-progress");

net = trainnet(augimdsTrain,net,"crossentropy",options);

% Inference — class names come from training data, not the pretrained net
classNames = categories(imdsTrain.Labels);
scores = minibatchpredict(net,imdsTest);
labels = scores2label(scores,classNames);

imagePretrainedNetwork returns class names only when both NumClasses and NumResponses are unset (pretrained mode, no transfer learning).

Workflow: Training

Check the Decision section above first — tabular data goes to fitrnet/fitcnet unless you need a non-LBFGS solver or a non-MSE/cross-entropy loss.

Standard training

% Define network
numChannels = 3;
numClasses = 5;
layers = [
    sequenceInputLayer(numChannels,Normalization="zscore")
    lstmLayer(100,OutputMode="last")
    fullyConnectedLayer(numClasses)
    softmaxLayer];

% Training options
options = trainingOptions("adam", ...
    MaxEpochs=30, ...
    MiniBatchSize=128, ...
    ValidationData={XVal,TVal}, ...
    Metrics="accuracy", ...
    Plots="training-progress");

% Train
net = trainnet(XTrain,TTrain,layers,"crossentropy",options);

Always normalize inputs. Set Normalization on the input layer (see example above). For regression, also normalize targets:

• R2026a+: append inverseNormalizationLayer to the last layer and set NormalizeTargets=true in trainingOptions
• Pre-R2026a: manually z-score targets before training and denormalize predictions at inference

See references/normalization.md for both workflows.

Custom loss function for multi-output

The function handle receives network outputs then targets, in order. Pass categorical targets directly — trainnet encodes them automatically.

lossFcn = @(Y1,Y2,T1,T2) crossentropy(Y1,T1) + mse(Y2,T2);

net = trainnet(ds,net,lossFcn,options);

For the full multi-output recipe (OutputNames alignment, combined datastores, testnet evaluation), see references/multi-output-training.md.

Since the multi-output loss is a function handle, it can be accelerated using dlaccelerate. Always inform the user that their custom loss function handle can be accelerated with dlaccelerate for faster training — even if they did not ask to speed up training. See references/dlaccelerate-trainnet-custom-loss.md for the verification and production workflow.

Workflow: Simulink Export

• dlnetwork (small, all layers supported): use exportNetworkToSimulink
• dlnetwork (large, or has layers unsupported by exportNetworkToSimulink): use the Predict block at library path deeplib/Predict
• fitcnet/fitrnet models: use the ClassificationNeuralNetwork Predict or RegressionNeuralNetwork Predict blocks from statsLibrary/

See references/simulink-export.md for details.

Workflow: Deep Learning Code Acceleration with dlaccelerate — Diagnose, Fix, or Apply

Does NOT apply to:

• Simulink export
• MEX/Coder compilation
• fitcnet/fitrnet/nlarx workflows

Applies when:

• User has a custom training loop (dlfeval/dlgradient)
• User has a function that supports dlarray input and is long running
• User calls trainnet with a custom loss function handle (R2026a+)

Always attempt to dlaccelerate functions that:

• Are long-running.
• Have dlarray objects, structures of dlarray objects, or dlnetwork objects as inputs.
• Only perform operations that affect the output variables. Operations that do not affect the output variables include plotting, displaying output, and writing to files.

Check references/dlaccelerate-workflow.md Step 0 to determine whether the function is acceleratable before recommending.

Follow references/dlaccelerate-workflow.md for the full diagnostic/fix/improve process. Entry points:

• User says "make faster" or "speed up" and the code has a custom training loop: start at Step 0.
• Code already uses dlaccelerate with problems: start at Step 1 (identify antipatterns, apply fixes, verify).
• Custom training loop without dlaccelerate: start at Step 0 (requirements check, wrap, verify).
• Any function with dlarray input that is called repeatedly: start at Step 0. This includes custom model functions used for prediction.
• Custom inference function called repeatedly: same Step 0 applies. dlaccelerate is not training-specific — any repeatedly-called dlarray function benefits. Use minibatchpredict when the loop only calls predict with no custom pre- or postprocessing; use dlaccelerate when custom operations surround the predict call.
• trainnet + custom loss function handle (R2026a+): wrap the loss with dlaccelerate and pass the AcceleratedFunction to trainnet. See references/dlaccelerate-trainnet-custom-loss.md.

For antipatterns that break tracing and their fixes, see references/dlaccelerate-antipatterns.md.

dlaccelerate References

• references/dlaccelerate-workflow.md — full diagnostic/fix/improve process
• references/dlaccelerate-antipatterns.md — pattern catalog with BAD/GOOD examples
• references/dlaccelerate-custom-training-loop.md — acceleration levels, L2, clipping
• references/dlaccelerate-trainnet-custom-loss.md — trainnet + custom loss (R2026a+)
• references/dlaccelerate-variable-length-sequences.md — padding/bucketing strategies
• references/dlaccelerate-custom-layers.md — Acceleratable custom layers (nnet.layer.Acceleratable)
• references/dlaccelerate-measure-speedup.md — benchmarking methodology

Key Functions

Common Mistakes

Copyright 2026 The MathWorks, Inc.