Error when using 'useGPU' for training neural networks

I am receiving an error whenever I try to perform patternnet training with the 'useGPU' option. I am using MATLAB R2016a under Windows 7 SP1. The error also occurs in MATLAB R2015a under Windows 7 SP1. I have the latest NVIDIA CUDA toolkit installed (version 7.5). I also have the latest NVIDIA drivers installed for my GTX 780.

Whenever I run:

net=patternnet([10,10]);
net.trainFcn='trainlm';
net.layers{3}.transferFcn='softmax';
[net,tr]=train(net,x1',d1','useGPU','yes');

MATLAB defaults to using 'trainscg' for the default training method. The training is successfully performed using the GPU. Whenever I run:

net=patternnet([10,10]);
net.trainFcn='trainscg';
net.layers{3}.transferFcn='softmax';
[net,tr]=train(net,x1',d1','useGPU','yes');

MATLAB reports that there is an error using gpuArray/arrayfun:

 Error using gpuArray/arrayfun
        Variable 't' changed type. Consider renaming variable on left hand side of assignment.
        For more information see Tips and Restrictions.
        Error in 'netHints' (line: 666)

 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 153)
        [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 147)
        [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);

If trainFnc is set to anything other than 'trainlm', such as 'trainbfg', I receive the same error.

Oddly enough, I don't encounter this error when training a feedforwardnet. I was able to succesfully use 'trainbfg', for example, without any issues. My sense is that there might be some toolbox coding issue for patternnets.

Answer 1

[Amanjit Dulai]

I suspect that the problem here is related to the type for the variable d1. I managed to replicate your error by using a single precision array for d1. Try casting it to double precision.

[HyeongHun LEE]

I had a similar problem with this issue(for loop statement of gpu used-autoencoder), and solved with changed the array variable type(from single to double)in gpuarray. Thank you.