When using GPU with neural net, I run out of shared memory per block; is there a way to handle?

I want to train a neural net with several hundred images (75x75 pixels, or 5625 elements each). This works in native Matlab. When I try to train using 'useGPU' I get the error "The shared memory size for a kernel must be a positive integer, and must not exceed the device's limit on the amount of shared memory per block (49152 bytes)." coming from nnGPU.codeHints. The code:

net1=feedforwardnet(10);
xg=nndata2gpu(inputMatrix);
tg=nndata2gpu(targetMatrix);
net2=configure(net1,inputMatrix,targetMatrix);
net2=train(net2,xg,tg);

Is there a way to tell the neural net training system to process the training in smaller chunks? Or some other smarter way to do this?

Answer 1

[Mark Hudson Beale]

I was able to reproduce your error. In MATLAB 13a the nndata2gpu array transformation is no longer required and if gpuArray is used (instead of nndata2gpu) the required amount of shared memory will be reduced.

d = gpuDevice
d.MaxShmemPerBlock

Using 13a and gpuArray I was able to train the following random problem on a mobile GPU with these specs: NVIDIA GeForce GT 650M 1024 MB in MATLAB 13a.

x = rand(5626,500);
t = rand(1,500);
X = gpuArray(x);
T = gpuArray(t);
net = feedforwardnet(10);
net = configure(net,x,t);
net.trainFcn = 'trainscg';
net = train(net,X,T);

I hope that helps!

[William Engelke]

With this code (which looks right, by the way), I get: Error using network/train (line 293) Number of samples (rows of gpuArrays) of data arguments do not match.

Error in GPUNET (line 11) net = train(net,X,T);

When I look at the sizes of the inputs, they look right, as follows:

>> size(X)

ans =

        5626         500

>> size(T)

ans =

     1   500

Maybe the problem is that I am using Matlab 2012b - (?) perhaps some bug was fixed in the newer release... anyway, I have decided to approach the problem a different way, such that it does not require so many rows.