## Run Custom Training Loops on a GPU and in Parallel

You can speed up your custom training loops by running on a GPU, in parallel using multiple GPUs, or on a cluster.

It is recommended to train using a GPU or multiple GPUs. Only use single CPU or multiple CPUs if you do not have a GPU. CPUs are normally much slower that GPUs for both training and inference. Running on a single GPU typically offers much better performance than running on multiple CPU cores.

Note

This topic shows you how to perform custom training on GPUs, in parallel, and on the cloud. To learn about parallel and GPU workflows using the trainNetwork function, see:

Using a GPU or parallel options requires Parallel Computing Toolbox™. Using a GPU also requires a supported GPU device. For information on supported devices, see GPU Computing Requirements (Parallel Computing Toolbox). Using a remote cluster also requires MATLAB® Parallel Server™.

### Train Network on GPU

By default, custom training loops run on the CPU. Automatic differentiation using dlgradient and dlfeval supports running on the GPU when your data is on the GPU. To run a custom training loop on a GPU, simply convert your data to gpuArray (Parallel Computing Toolbox) during training.

You can use minibatchqueue to manage your data during training. minibatchqueue automatically prepares data for training, including custom preprocessing and converting data to dlarray and gpuArray. By default, minibatchqueue returns all mini-batch variables on the GPU if one is available. You can choose which variables to return on the GPU using the OutputEnvironment property.

For an example showing how to use minibatchqueue to train on the GPU, see Train Network Using Custom Training Loop.

Alternatively, you can manually convert your data to gpuArray within the training loop.

To easily specify the execution environment, create the variable executionEnvironment that 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 a gpuArray if necessary. The canUseGPU function checks for useable GPUs.

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

### Train Single Network in Parallel

When you train in parallel, each worker trains the network simultaneously using a portion of a mini-batch. This means that you must combine the gradients, loss, and any state parameters after each iteration, according to the proportion of the mini-batch processed by each worker.

You can train in parallel on your local machine, or on a remote cluster, for example, in the cloud. Start a parallel pool in the desired resources and partition your data between the workers. During training, combine the gradients, loss, and state after each iteration so that the learnable parameters on each worker update in synchronization. For an example showing how to perform custom training in parallel, see Train Network in Parallel with Custom Training Loop

#### Set Up Parallel Environment

It is recommended to train using a GPU or multiple GPUs. Only use single CPU or multiple CPUs if you do not have a GPU. CPUs are normally much slower that GPUs for both training and inference. Running on a single GPU typically offers much better performance than running on multiple CPU cores.

Set up the parallel environment that you want to use before training. Start a parallel pool using your desired resources. For training using multiple GPUs, start a parallel pool with as many workers as available GPUs. For best performance, MATLAB automatically assigns a different GPU to each worker.

If you are using your local machine, you can use canUseGPU and gpuDeviceCount (Parallel Computing Toolbox) to determine if you have GPUs available. For example, to check availabilities of GPUs and start a parallel pool with as many workers as available GPUs, use the following code:

if canUseGPU
executionEnvironment = "gpu";
numberOfGPUs = gpuDeviceCount("available");
pool = parpool(numberOfGPUs);
else
executionEnvironment = "cpu";
pool = parpool;
end

If you are running using a remote cluster, for example, a cluster in the cloud, start a parallel pool with as many workers as the number of GPUs per machine multiplied by the number of machines.

For more information on selecting specific GPUs, see Select Particular GPUs to Use for Training.

#### Specify Mini-Batch Size and Partition Data

Specify the mini-batch size that you want to use during training. For GPU training, a recommended practice is to scale up the mini-batch size linearly with the number of GPUs, in order to keep the workload on each GPU constant. For example, if you are training on a single GPU using a mini-batch size of 64, and you want to scale up to training with four GPUs of the same type, you can increase the mini-batch size to 256 so that each GPU processes 64 observations per iteration.

You can use the following code to scale up the mini-batch size by the number of workers, where N is the number of workers in your parallel pool.

if executionEnvironment == "gpu"
miniBatchSize = miniBatchSize .* N
end

If you want to use a mini-batch size that not exactly divisible by the number of workers in your parallel pool, then distribute the remainder across the workers.

workerMiniBatchSize = floor(miniBatchSize ./ repmat(N,1,N));
remainder = miniBatchSize - sum(workerMiniBatchSize);
workerMiniBatchSize = workerMiniBatchSize + [ones(1,remainder) zeros(1,N-remainder)]

At the start of training, shuffle your data. Partition your data so that each worker has access to a portion of the mini-batch. To partition a datastore, use the partition function.

You can use minibatchqueue to manage the data on each worker during training. minibatchqueue automatically prepares data for training, including custom preprocessing and converting data to dlarray and gpuArray. Create a minibatchqueue on each worker using the partitioned datastore. Set the MiniBatchSize property using the mini-batch sizes calculated for each worker.

At the start of each training iteration, use the gop (Parallel Computing Toolbox) function to check that all worker minibatchqueue objects can return data. If any worker runs out of data, training stops. If your overall mini-batch size is not exactly divisible by the number of workers and you do not discard partial mini-batches, some workers might run out of data before others.

Write your training code inside an spmd (Parallel Computing Toolbox) block, so that the training loop executes on each worker.

spmd
% Reset and shuffle the datastore.
reset(augimdsTrain);
augimdsTrain = shuffle(augimdsTrain);

% Partition datastore.
workerImds = partition(augimdsTrain,N,spmdIndex);

% Create minibatchqueue using partitioned datastore on each worker
workerMbq = minibatchqueue(workerImds,...
"MiniBatchSize",workerMiniBatchSize(spmdIndex),...
"MiniBatchFcn",@preprocessMiniBatch);

...

for epoch = 1:numEpochs

% Reset and shuffle minibatchqueue on each worker.
shuffle(workerMbq);

% Loop over mini-batches.
while gop(@and,hasdata(workerMbq))

% Custom training loop
...

end
...
end
end

To ensure that the network on each worker learns from all data and not just the data on that worker, aggregate the gradients and use the aggregated gradients to update the network on each worker.

For example, suppose you are training the network net, using the model loss function modelLoss. Your training loop contains the following code for evaluating the loss, gradients, and statistics on each worker:

workerX and workerT are the predictor and true response on each worker, respectively.

To aggregate the gradients, use a weighted sum. Define a helper function to sum the gradients.

end

Inside the training loop, use dlupdate to apply the function to the gradients of each learnable parameter.

#### Aggregate Loss and Accuracy

To find the network loss and accuracy, for example, to plot them during training to monitor training progress, aggregate the values of the loss and accuracy on all of the workers. Typically, the aggregated value is the sum of the value on each worker, weighted by the proportion of the mini-batch used on each worker. To aggregate the losses and accuracy each iteration, calculate the weight factor for each worker and use gplus (Parallel Computing Toolbox) to sum the values on each worker.

workerNormalizationFactor = workerMiniBatchSize(spmdIndex)./miniBatchSize;
loss = gplus(workerNormalizationFactor*extractdata(dlworkerLoss));
accuracy = gplus(workerNormalizationFactor*extractdata(dlworkerAccuracy));

#### Aggregate Statistics

If your network contains layers that track the statistics of your training data, such as batch normalization layers, then you must aggregate the statistics across all workers after each training iteration. Doing so ensures that the network learns statistics that are representative of the entire training set.

You can identify the layers that contain statistics information before training. For example, if you are using a dlnetwork with batch normalization layers, you can use the following code to find the relevant layers.

batchNormLayers = arrayfun(@(l)isa(l,'nnet.cnn.layer.BatchNormalizationLayer'),net.Layers);
batchNormLayersNames = string({net.Layers(batchNormLayers).Name});
state = net.State;
isBatchNormalizationStateMean = ismember(state.Layer,batchNormLayersNames) & state.Parameter == "TrainedMean";
isBatchNormalizationStateVariance = ismember(state.Layer,batchNormLayersNames) & state.Parameter == "TrainedVariance";
Define a helper function to aggregate the statistics you are using. Batch normalization layers track the mean and variance of the input data. You can aggregate the mean on all the workers using a weighted average. To calculate the aggregated variance ${s}_{c}^{2}$, use a formula of the following form.

${s}_{c}^{2}=\frac{1}{M}\sum _{j=1}^{N}{m}_{j}\left({s}_{j}^{2}+{\left({\overline{x}}_{j}-{\overline{x}}_{c}\right)}^{2}\right)$

N is the total number of workers, M is the total number of observations in a mini-batch, mj is the number of observations processed on the jth worker, ${\overline{x}}_{j}$ and ${s}_{j}^{2}$ are the mean and variance statistics calculated on that worker, and ${\overline{x}}_{c}$ is the aggregated mean across all workers.

function state = aggregateState(state,factor,...
isBatchNormalizationStateMean,isBatchNormalizationStateVariance)

stateMeans = state.Value(isBatchNormalizationStateMean);
stateVariances = state.Value(isBatchNormalizationStateVariance);

for j = 1:numel(stateMeans)
meanVal = stateMeans{j};
varVal = stateVariances{j};

% Calculate combined mean
combinedMean = gplus(factor*meanVal);

% Calculate combined variance terms to sum
varTerm = factor.*(varVal + (meanVal - combinedMean).^2);

% Update state
stateMeans{j} = combinedMean;
stateVariances{j} = gplus(varTerm);
end

state.Value(isBatchNormalizationStateMean) = stateMeans;
state.Value(isBatchNormalizationStateVariance) = stateVariances;
end

Inside the training loop, use the helper function to update the state of the batch normalization layers with the combined mean and variance.

net.State = aggregateState(workerState,workerNormalizationFactor,...
isBatchNormalizationStateMean,isBatchNormalizationStateVariance);

#### Plot Results During Training

If you want to plot results during training, you can send data from the workers to the client using a DataQueue object.

To easily specify that the plot should be on or off, create the variable plots that contains either "training-progress" or "none".

plots = "training-progress";

Before training, initialize the DataQueue and the animated line using the animatedline function.

if plots == "training-progress"
figure
lineLossTrain = animatedline('Color',[0.85 0.325 0.098]);
ylim([0 inf])
xlabel("Iteration")
ylabel("Loss")
grid on
end
Create the DataQueue object. Use afterEach to call the helper function displayTrainingProgress each time data is sent from the worker to the client.
Q = parallel.pool.DataQueue;
displayFcn = @(x) displayTrainingProgress(x,lineLossTrain);
afterEach(Q,displayFcn);
The displayTrainingProgress helper function contains the code used to add points to the animated line and display the training epoch and duration.
function displayTrainingProgress (data,line)
D = duration(0,0,data(4),'Format','hh:mm:ss');
title("Epoch: " + data(1) + ", Elapsed: " + string(D))
drawnow
end

Inside the training loop, at the end of each epoch, use the DataQueue to send the training data from the workers to the client. At the end of each iteration, the aggregated loss is the same on each worker, so you can send data from a single worker.

% Display training progress information.
if spmdIndex == 1
data = [epoch loss iteration toc(start)];
send(Q,gather(data));
end

### Train Multiple Networks in Parallel

To train multiple networks in parallel, start a parallel pool in your desired resources and use parfor (Parallel Computing Toolbox) to train a single network on each worker.

You can run locally or using a remote cluster. Using a remote cluster requires MATLAB Parallel Server. For more information about managing cluster resources, see Discover Clusters and Use Cluster Profiles (Parallel Computing Toolbox). If you have multiple GPUs and want to exclude some from training, you can choose the GPUs you use to train on. For more information on selecting specific GPUs, see Select Particular GPUs to Use for Training.

You can modify the network or training parameters on each worker to perform parameter sweeps in parallel. For example, in networks is an array of dlnetwork objects, you can use code of the following form to train multiple different networks using the same data.

parpool ("Processes",numNetworks);

parfor idx = 1:numNetworks
iteration = 0;
velocity = [];

% Allocate one network per worker
net = networks(idx)

% Loop over epochs.
for epoch = 1:numEpochs
% Shuffle data.
shuffle(mbq);

% Loop over mini-batches.
while hasdata(mbq)
iteration = iteration + 1;

% Custom training loop
...

end

end

% Send the trained networks back to the client.
trainedNetworks{idx} = net;
end
After parfor finishes, trainedNetworks contains the resulting networks trained by the workers.

#### Plot Results During Training

To monitor training progress on the workers, you can use a DataQueue to send data back from the workers.

To easily specify that the plot should be on or off, create the variable plots that contains either "training-progress" or "none".

plots = "training-progress";

Before training, initialize the DataQueue and the animated lines using the animatedline function. Create a subplot for each network you are training.

if plots == "training-progress"
f = figure;
f.Visible = true;
for i=1:numNetworks
subplot(numNetworks,1,i)
xlabel('Iteration');
ylabel('loss');
lines(i) = animatedline;
end
end
Create the DataQueue object. Use afterEach to call the helper function displayTrainingProgress each time data is sent from the worker to the client.
Q = parallel.pool.DataQueue;
displayFcn = @(x) displayTrainingProgress(x,lines);
afterEach(Q,displayFcn);
The displayTrainingProgress helper function contains the code used to add points to the animated lines.
function displayTrainingProgress (data,lines)
D = duration(0,0,data(5),'Format','hh:mm:ss');
title("Epoch: " + data(2) + ", Elapsed: " + string(D))
drawnow limitrate nocallbacks
end

Inside the training loop, at the end of each iteration, use the DataQueue to send the training data from the workers to the client. Send the parfor loop index as well as the training information so that the points are added to the correct line for each worker.

% Display training progress information.
data = [idx epoch loss iteration toc(start)];
send(Q,gather(data));

### Use Experiment Manager to Train in Parallel

You can use Experiment Manager to run your custom training loops in parallel. You can either run multiple trails at the same time, or run a single trial at a time using parallel resources.

To run multiple trials at the same time using one parallel worker for each trial, set up your custom training experiment an enable the Use Parallel option before running your experiment.

To run a single trial at a time using multiple parallel workers, define your parallel environment in your experiment training function and use an spmd block to train the network in parallel. For more information on training a single network in parallel with a custom training loop, see Train Single Network in Parallel.

For more information on training in parallel using Experiment Manager, see Use Experiment Manager to Train in Parallel.