minibatchqueue or arrayDatastore drops my data precision from double to single

I get XTrain from MNIST by processImagesMNIST and put it on GPU, so its type is gpuArray dlarray.

then I use these code to make minibatches:

```

miniBatchSize = 128;

dsTrain = arrayDatastore(XTrain,IterationDimension=4);
% numOutputs = 1;

mbqTest = minibatchqueue(dsTrain,1, ...
    MiniBatchSize = miniBatchSize, ...
    MiniBatchFcn=@preprocessMiniBatch, ...
    MiniBatchFormat="SSCB", ...
    PartialMiniBatch="discard");

% numObservationsTrain = size(XTrain,4);
% numIterationsPerEpoch = ceil(numObservationsTrain / miniBatchSize);
% numIterations = numEpochs * numIterationsPerEpoch;

%% test batch order
i=0;
while hasdata(mbqTest)
    i = i+1;
    x = next(mbqTest);
    if ~hasdata(mbqTest)
        disp(i)
    end
end
```

And I find that x is single gpuArray dlarray, XTrain is gpuArray dlarray.

I wonder which part makes it lower the precision.

And how to avoid this?

[Runcong Kuang]

I find that minibatchqueue has a 'name,value' pair which is outputCast, 'single' as default. So I assign 'double' to it. But I haven't checked if the GPU training process does use single precision, as mentioned by @Walter Roberson.

But thanks for your input @Walter Roberson.

Answer 1

[Walter Roberson]

Gpu training does not support double precision. If you look at the available options, precision cannot be selected.