Using GPU with built-in machine learning functons

I wanted to check the speed-up of my classification task when using the GPU instead of the CPU. More specifically I tried something like:

          tic
          NB=fitcnb(gpuArray(train_values), train_labels,'KFold',5,'CrossVal','on');
          kfoldLoss(NB)
          time_upNB=toc;

but I am getting a "Conversion to double from gpuArray is not possible." error. Is my syntax wrong or is this not possible?

[Astarag Chattopadhyay]

Can you give more details about the data types of the variables that you are using in your code?

[Tasos]

Yes, of course. The array "train_values" contains double numbers, while the vector "train_labels" contains zeros and ones.

[Astarag Chattopadhyay]

Hi Tasos,

It is hard to tell from which part of the code you are getting this error. In general, this error occurs when you try to save any gpuArray calculation result to a variable. As an example

A = magic(5);
Agpu = gpuArray(A);
B = zeros(5);
for i = 1:5
  B(i,i) = Agpu(i,i) * Agpu(i,i);
end

This code snippet will throw the same error as you are getting. You need to preallocate B as gpuArray to workaround this:

iA = magic(5);
Agpu = gpuArray(A);
Bgpu = gpuArray(zeros(5));
for i = 1:5
  Bgpu(i,i) = Agpu(i,i) * Agpu(i,i);
end

I will suggest you put breakpoints in your code to see where you are getting the error and most probably you need do a gpuArray conversion at that line of code.