A working example of Neural Network "train" with GPU and inputs / targets as SINGLE matrices

Hi guys,

I'm starting to employ my CUDA GPU with the Neural Network toolbox, but if I feed the trainer with double matrices I soon run out of CUDA memory. As a solution to improve both computational efficiency and data size limit, I've thought about employing single matrices. The problem is that I always get this error, and I would like to kindly ask you for a working example of GPU training with single matrices. Thanks in advance!

Error using gpuArray/arrayfun
Variable 'y' changed type. Consider renaming variable on left hand side of assignment.
For more information see Tips and Restrictions.
 Error in 'netHints' (line: 667)

Error in nnGPUOp.bg (line 133)
        [perfs1,perfs2,perfs3,N1,N2,N3,dy] =
        arrayfun(hints.performFcns.perf_dy_3masks,ti,yi,en,ew,mask1,mask2,mask3,S);

Error in nnGPUOp.perfsGrad (line 5)
[gWB,Perfs,PerfN] = nnGPUOp.bg...

Error in nnCalcLib/perfsGrad (line 294)
                lib.calcMode.perfsGrad(calcNet,lib.calcData,lib.calcHints);

Error in trainscg>initializeTraining (line 155)
[worker.perf,worker.vperf,worker.tperf,worker.gWB,worker.gradient] = calcLib.perfsGrad(calcNet);

Error in nnet.train.trainNetwork>trainNetworkInMainThread (line 28)
worker = localFcns.initializeTraining(archNet,calcLib,calcNet,tr);

Error in nnet.train.trainNetwork (line 16)
    [archNet,tr] = trainNetworkInMainThread(archNet,rawData,calcLib,calcNet,tr,feedback,localFcns);

Error in trainscg>train_network (line 149)
[archNet,tr] = nnet.train.trainNetwork(archNet,rawData,calcLib,calcNet,tr,localfunctions);

Error in trainscg (line 59)
            [out1,out2] = train_network(varargin{2:end});

Error in network/train (line 369)
    [net,tr] = feval(trainFcn,'apply',net,data,calcLib,calcNet,tr);

Error in test (line 52)
net2      = train(net2,inputs_g,targets_g,'useGpu','yes');    % Execute on GPU with NNET formatted
gpuArrays

[Cameron Lee]

I'm not sure if this issue is stil persisting for you, but I also have the same problem. I've started another thread (here: https://www.mathworks.com/matlabcentral/answers/500655-training-nn-with-single-precision-data-on-gpu?s_tid=prof_contriblnk) that has gotten the attention of a Mathworks staff memeber. Perhaps he will be able to help.

I've seen others with this issue as well, and even one that had an answer bascially saying to use double-precision to get around the problem. However, for large datasets, single-precision data processed on most GPUs would likely be exponentially faster, so while it is a work-around for the issue, it doesn't seem to me to be the ideal solution. So - we'll see!

Answer 1

[Brad Lipovsky]

I have the same problem.