Can I use parallel computing to train a DNN?

Question

Callum Giblett 2019-2-14

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/444989-can-i-use-parallel-computing-to-train-a-dnn

回答： Joss Knight 2019-2-15

Contained below is my code for a Neural Network I have designed. The network training runs fine on a single CPU but it is very slow, is it possible to use a parallel set up? I have made what i think are the required modifications but i get an error (also contained below), what is meant by a recurrent network and is there any way to overcome this error?

Error using trainNetwork (line 150)

Parallel training of recurrent networks is not supported. 'ExecutionEnvironment' value in trainingOptions function must be 'auto', 'gpu', or 'cpu'.

Error in Network2 (line 49)

network2 = trainNetwork(TrainingInputData,TrainingOutputData,layers,options)

Caused by:

Error using nnet.internal.cnn.assembler.setupExecutionEnvironment (line 17)

Parallel training of recurrent networks is not supported. 'ExecutionEnvironment' value in trainingOptions function must be 'auto', 'gpu', or

'cpu'.

clc
clear
parpool
[outputdata, inputdata,specifications] = dataprep();
TrainingOutputData = outputdata;
TrainingInputData = inputdata;
TestingOutputData = cell(1,1);
TestingInputData = cell(1,1);
TestingOutputData{1,1} = outputdata{4,1};
TestingInputData{1,1} = inputdata{4,1};
TrainingOutputData(4,:)=[];
TrainingInputData(4,:)=[];
maxEpochs = 10;
miniBatchSize = 1;
Neurons = 150000;
numResponses = size(TrainingOutputData{1},1);
featureDimension = size(TrainingInputData{1},1);
layers = [ ...
    sequenceInputLayer(featureDimension)
    fullyConnectedLayer(Neurons)
    fullyConnectedLayer(numResponses)
    regressionLayer];
options = trainingOptions('sgdm', ...
    'MaxEpochs',maxEpochs, ...
    'MiniBatchSize',miniBatchSize, ...
    'InitialLearnRate',0.01, ...
    'LearnRateSchedule','piecewise',...
    'LearnRateDropPeriod',3,...
    'LearnRateDropFactor',0.1,...
    'GradientThreshold',1, ...
    'Shuffle','never', ...
    'Plots','training-progress',...
    'ValidationData',{TestingInputData,TestingOutputData},...
    'ValidationFrequency', 1,...
    'Verbose', 1,...
    'VerboseFrequency', 1,...
    'ExecutionEnvironment', 'parallel');
network2 = trainNetwork(TrainingInputData,TrainingOutputData,layers,options)