Main Content

train

Train shallow neural network

Description

This function trains a shallow neural network. For training deep learning networks (such as convolutional or LSTM networks), use the trainnet function.

trainedNet = train(net,X,T,Xi,Ai,EW) trains a network net according to net.trainFcn and net.trainParam.

example

[trainedNet,tr] = train(net,X,T,Xi,Ai,EW) also returns a training record.

[trainedNet,tr] = train(net,X,T,Xi,Ai,EW,Name,Value) trains a network with additional options specified by one or more name-value pair arguments.

example

Examples

collapse all

Here input x and targets t define a simple function that you can plot:

x = [0 1 2 3 4 5 6 7 8];
t = [0 0.84 0.91 0.14 -0.77 -0.96 -0.28 0.66 0.99];
plot(x,t,'o')

Here feedforwardnet creates a two-layer feed-forward network. The network has one hidden layer with ten neurons.

net = feedforwardnet(10);
net = configure(net,x,t);
y1 = net(x)
plot(x,t,'o',x,y1,'x')

The network is trained and then resimulated.

net = train(net,x,t);
y2 = net(x)
plot(x,t,'o',x,y1,'x',x,y2,'*')

This example trains an open-loop nonlinear-autoregressive network with external input, to model a levitated magnet system defined by a control current x and the magnet’s vertical position response t, then simulates the network. The function preparets prepares the data before training and simulation. It creates the open-loop network’s combined inputs xo, which contains both the external input x and previous values of position t. It also prepares the delay states xi.

[x,t] = maglev_dataset;
net = narxnet(10);
[xo,xi,~,to] = preparets(net,x,{},t);
net = train(net,xo,to,xi);
y = net(xo,xi)

This same system can also be simulated in closed-loop form.

netc = closeloop(net);
view(netc)
[xc,xi,ai,tc] = preparets(netc,x,{},t);
yc = netc(xc,xi,ai);

If you have hardware resources that are not being used during training, you might be able to train your network faster and use a dataset that otherwise would not fit in memory by training in parallel. Training a network in parallel requires Parallel Computing Toolbox™ and is supported for backpropagation training only, not for self-organizing maps.

Here training and simulation happens across parallel MATLAB workers.

[X,T] = vinyl_dataset;
net = feedforwardnet(10);
net = train(net,X,T,'useParallel','yes','showResources','yes');
Y = net(X);

Use Composite values to distribute the data manually, and get back the results as a Composite value. If the data is loaded as it is distributed then while each piece of the dataset must fit in RAM, the entire dataset is limited only by the total RAM of all the workers.

[X,T] = vinyl_dataset;
Q = size(X,2);
Xc = Composite;
Tc = Composite;
numWorkers = numel(Xc);
ind = [0 ceil((1:numWorkers)*(Q/numWorkers))];
for i=1:numWorkers
    indi = (ind(i)+1):ind(i+1);
    Xc{i} = X(:,indi);
    Tc{i} = T(:,indi);
end
net = feedforwardnet;
net = configure(net,X,T);
net = train(net,Xc,Tc);
Yc = net(Xc);

Note in the example above the configure function was used to set the dimensions and processing settings of the network's inputs. This normally happens automatically when train is called, but when providing composite data this step must be done manually with non-Composite data.

Networks can be trained using the current GPU device, if it is supported by Parallel Computing Toolbox. GPU training is currently supported for backpropagation training only, not for self-organizing maps.

[X,T] = vinyl_dataset;
net = feedforwardnet(10);
net = train(net,X,T,'useGPU','yes');
y = net(X); 

To put the data on a GPU manually:

[X,T] = vinyl_dataset;
Xgpu = gpuArray(X);
Tgpu = gpuArray(T);
net = configure(net,X,T);
net = train(net,Xgpu,Tgpu);
Ygpu = net(Xgpu);
Y = gather(Ygpu); 

Note in the example above the function configure was used to set the dimensions and processing settings of the network's inputs. This normally happens automatically when train is called, but when providing gpuArray data this step must be done manually with non-gpuArray data.

To run in parallel, with workers each assigned to a different unique GPU, with extra workers running on CPU:

net = train(net,X,T,'useParallel','yes','useGPU','yes');
y = net(X);

Using only workers with unique GPUs might result in higher speed, as CPU workers might not keep up.

net = train(net,X,T,'useParallel','yes','useGPU','only');
Y = net(X);

Here a network is trained with checkpoints saved at a rate no greater than once every two minutes.

[x,t] = vinyl_dataset;
net = fitnet([60 30]);
net = train(net,x,t,'CheckpointFile','MyCheckpoint','CheckpointDelay',120);

After a computer failure, the latest network can be recovered and used to continue training from the point of failure. The checkpoint file includes a structure variable checkpoint, which includes the network, training record, filename, time, and number.

[x,t] = vinyl_dataset;
load MyCheckpoint
net = checkpoint.net;
net = train(net,x,t,'CheckpointFile','MyCheckpoint');

Input Arguments

collapse all

Input network, specified as a network object. To create a network object, use for example, feedforwardnet or narxnet.

Network inputs, specified as an R-by-Q matrix or an Ni-by-TS cell array, where

  • R is the input size

  • Q is the batch size

  • Ni = net.numInputs

  • TS is the number of time steps

train arguments can have two formats: matrices, for static problems and networks with single inputs and outputs, and cell arrays for multiple timesteps and networks with multiple inputs and outputs.

  • The matrix format can be used if only one time step is to be simulated (TS = 1). It is convenient for networks with only one input and output, but can be used with networks that have more. When the network has multiple inputs, the matrix size is (sum of Ri)-by-Q.

  • The cell array format is more general, and more convenient for networks with multiple inputs and outputs, allowing sequences of inputs to be presented. Each element X{i,ts} is an Ri-by-Q matrix, where Ri = net.inputs{i}.size.

If Composite data is used, then 'useParallel' is automatically set to 'yes'. The function takes Composite data and returns Composite results.

If gpuArray data is used, then 'useGPU' is automatically set to 'yes'. The function takes gpuArray data and returns gpuArray results

Note

If a column of X contains at least one NaN, train does not use that column for training, testing, or validation.

Network targets, specified as a U-by-Q matrix or an No-by-TS cell array, where

  • U is the output size

  • Q is the batch size

  • No = net.numOutputs

  • TS is the number of time steps

train arguments can have two formats: matrices, for static problems and networks with single inputs and outputs, and cell arrays for multiple timesteps and networks with multiple inputs and outputs.

  • The matrix format can be used if only one time step is to be simulated (TS = 1). It is convenient for networks with only one input and output, but can be used with networks that have more. When the network has multiple inputs, the matrix size is (sum of Ui)-by-Q.

  • The cell array format is more general, and more convenient for networks with multiple inputs and outputs, allowing sequences of inputs to be presented. Each element T{i,ts} is a Ui-by-Q matrix, where Ui = net.outputs{i}.size.

If Composite data is used, then 'useParallel' is automatically set to 'yes'. The function takes Composite data and returns Composite results.

If gpuArray data is used, then 'useGPU' is automatically set to 'yes'. The function takes gpuArray data and returns gpuArray results

Note that T is optional and need only be used for networks that require targets.

Note

Any NaN values in the targets T are treated as missing data. If an element of T is NaN, that element is not used for training, testing, or validation.

Initial input delay conditions, specified as an Ni-by-ID cell array or an R-by-(ID*Q) matrix, where

  • ID = net.numInputDelays

  • Ni = net.numInputs

  • R is the input size

  • Q is the batch size

For cell array input, the columns of Xi are ordered from the oldest delay condition to the most recent: Xi{i,k} is the input i at time ts = k - ID.

Xi is also optional and need only be used for networks that have input or layer delays.

Initial layer delay conditions, specified as a Nl-by-LD cell array or a (sum of Si)-by-(LD*Q) matrix, where

  • Nl = net.numLayers

  • LD = net.numLayerDelays

  • Si = net.layers{i}.size

  • Q is the batch size

For cell array input, the columns of Ai are ordered from the oldest delay condition to the most recent: Ai{i,k} is the layer output i at time ts = k - LD.

Error weights, specified as a No-by-TS cell array or a (sum of Ui)-by-Q matrix, where

  • No = net.numOutputs

  • TS is the number of time steps

  • Ui = net.outputs{i}.size

  • Q is the batch size

For cell array input. each element EW{i,ts} is a Ui-by-Q matrix, where

  • Ui = net.outputs{i}.size

  • Q is the batch size

The error weights EW can also have a size of 1 in place of all or any of No, TS, Ui or Q. In that case, EW is automatically dimension extended to match the targets T. This allows for conveniently weighting the importance in any dimension (such as per sample) while having equal importance across another (such as time, with TS=1). If all dimensions are 1, for instance if EW = {1}, then all target values are treated with the same importance. That is the default value of EW.

As noted above, the error weights EW can be of the same dimensions as the targets T, or have some dimensions set to 1. For instance if EW is 1-by-Q, then target samples will have different importances, but each element in a sample will have the same importance. If EW is (sum of Ui)-by-1, then each output element has a different importance, with all samples treated with the same importance.

Name-Value Arguments

Specify optional pairs of arguments as Name1=Value1,...,NameN=ValueN, where Name is the argument name and Value is the corresponding value. Name-value arguments must appear after other arguments, but the order of the pairs does not matter.

Before R2021a, use commas to separate each name and value, and enclose Name in quotes.

Example: 'useParallel','yes'

Option to specify parallel calculations, specified as 'yes' or 'no'.

  • 'no' – Calculations occur on normal MATLAB thread. This is the default 'useParallel' setting.

  • 'yes' – Calculations occur on parallel workers if a parallel pool is open. If there is no parallel pool open, the software starts one using the default cluster profile.

Option to specify GPU calculations, specified as 'yes', 'no', or 'only'.

  • 'no' – Calculations occur on the CPU. This is the default 'useGPU' setting.

  • 'yes' – Calculations occur on the current gpuDevice if it is a supported GPU (See Parallel Computing Toolbox for GPU requirements.) If the current gpuDevice is not supported, calculations remain on the CPU. If 'useParallel' is also 'yes' and a parallel pool is open, then each worker with a unique GPU uses that GPU, other workers run calculations on their respective CPU cores.

  • 'only' – If no parallel pool is open, then this setting is the same as 'yes'. If a parallel pool is open then only workers with unique GPUs are used. However, if a parallel pool is open, but no supported GPUs are available, then calculations revert to performing on all worker CPUs.

Option to show resources, specified as 'yes' or 'no'.

  • 'no' – Do not display computing resources used at the command line. This is the default setting.

  • 'yes' – Show at the command line a summary of the computing resources actually used. The actual resources may differ from the requested resources, if parallel or GPU computing is requested but a parallel pool is not open or a supported GPU is not available. When parallel workers are used, each worker’s computation mode is described, including workers in the pool that are not used.

Memory reduction, specified as a positive integer.

For most neural networks, the default CPU training computation mode is a compiled MEX algorithm. However, for large networks the calculations might occur with a MATLAB® calculation mode. This can be confirmed using 'showResources'. If MATLAB is being used and memory is an issue, setting the reduction option to a value N greater than 1, reduces much of the temporary storage required to train by a factor of N, in exchange for longer training times.

Checkpoint file, specified as a character vector.

The value for 'CheckpointFile' can be set to a filename to save in the current working folder, to a file path in another folder, or to an empty string to disable checkpoint saves (the default value).

Checkpoint delay, specified as a nonnegative integer.

The optional parameter 'CheckpointDelay' limits how often saves happen. Limiting the frequency of checkpoints can improve efficiency by keeping the amount of time saving checkpoints low compared to the time spent in calculations. It has a default value of 60, which means that checkpoint saves do not happen more than once per minute. Set the value of 'CheckpointDelay' to 0 if you want checkpoint saves to occur only once every epoch.

Output Arguments

collapse all

Trained network, returned as a network object.

Training record (epoch and perf), returned as a structure whose fields depend on the network training function (net.NET.trainFcn). It can include fields such as:

  • Training, data division, and performance functions and parameters

  • Data division indices for training, validation and test sets

  • Data division masks for training validation and test sets

  • Number of epochs (num_epochs) and the best epoch (best_epoch)

  • A list of training state names (states)

  • Fields for each state name recording its value throughout training

  • Best performances of the network, evaluated at each epoch: best performance on the training set (best_perf), best performance on the validation set (best_vperf), and best performance on the test set (best_tperf)

Algorithms

train calls the function indicated by net.trainFcn, using the training parameter values indicated by net.trainParam.

Typically one epoch of training is defined as a single presentation of all input vectors to the network. The network is then updated according to the results of all those presentations.

Training occurs until a maximum number of epochs occurs, the performance goal is met, or any other stopping condition of the function net.trainFcn occurs.

Some training functions depart from this norm by presenting only one input vector (or sequence) each epoch. An input vector (or sequence) is chosen randomly for each epoch from concurrent input vectors (or sequences). competlayer returns networks that use trainru, a training function that does this.

Version History

Introduced before R2006a

See Also

| | |