trainNetwork — Train a supported in-memory feed-forward network.

trainNetwork(X,Y,layers,options) trains a supported sequential feed-forward Deep Learning network over in-memory numeric or logical predictor data.

Syntax

net = trainNetwork(X, Y, layers, options)
[net, info] = trainNetwork(X, Y, layers, options)

How trainNetwork works

• Supports the four-argument form with 2-D predictor matrices where rows are observations and columns are features.
• Layer collections may be layer objects, cell arrays of layer objects, object lists, or simple layerGraph metadata objects.
• Supported trainable topology is the same sequential feed-forward subset used by forward: featureInputLayer, fullyConnectedLayer, reluLayer, eluLayer, softmaxLayer, classificationLayer, and regressionLayer.
• For regression, Y must be a numeric or logical matrix with one row per observation and the same output width as the network.
• For classification, Y may be numeric integer labels, string labels, categorical labels, or a one-hot/probability matrix. The returned trained network stores sorted class labels in Classes when labels are provided.
• trainingOptions metadata controls sgdm and adam training, including MaxEpochs, MiniBatchSize, InitialLearnRate, Momentum, Adam decay factors, Epsilon, deterministic Shuffle handling, and CPU/auto ExecutionEnvironment values.
• The returned object is a SeriesNetwork for layer arrays and a DAGNetwork for simple layerGraph inputs. It includes updated layer weights, Learnables, TrainingInfo, TrainingLoss, and TrainedWith metadata.
• Datastores, image/sequence/recurrent/convolutional/normalization/custom layers, branching DAG execution, validation data, callbacks, checkpoints, plots, multi-GPU training, and automatic differentiation tracing remain outside this slice and raise explicit compatibility errors or are stored as inert options.

GPU memory and residency

trainNetwork currently gathers gpuArray inputs to host before training when the execution environment is auto or cpu. Explicit ExecutionEnvironment values of gpu, multi-gpu, or parallel raise a compatibility error because native GPU training kernels and provider-resident Deep Learning layer execution remain future acceleration work.

Examples

Train a small regression network

layers = {featureInputLayer(1,'Name','x'); fullyConnectedLayer(1,'Name','fc'); regressionLayer('Name','out')};
opts = trainingOptions('adam','MaxEpochs',120,'MiniBatchSize',4,'InitialLearnRate',0.05,'Shuffle','never');
net = trainNetwork([0;1;2;3], [1;3;5;7], layers, opts)

Expected output:

net is a trained SeriesNetwork compatibility object usable with predict.

Train a two-class network

layers = {featureInputLayer(2,'Name','in'); fullyConnectedLayer(2,'Name','fc'); softmaxLayer('Name','prob'); classificationLayer('Name','class')};
net = trainNetwork(X, labels, layers, opts)

Expected output:

net.Classes records the sorted class labels.

Using trainNetwork with coding agents

Open a RunMat example with live inputs, then ask the agent to explain how trainNetwork changes the result.

Related Deep Learning functions

About RunMat

RunMat is an open-source runtime that executes MATLAB-syntax code blazing on any GPU. It is licensed under the Apache 2.0 license.

• RunMat automatically optimizes your math for GPU execution on Apple, Nvidia, and AMD hardware. No code changes needed. Simulations that took hours now take minutes.
• Start running code in seconds. RunMat runs in the browser, on the desktop, or from the CLI. No license server, no IT ticket.

Getting started · Benchmarks · Pricing