search

vlfeat / matconvnet

MatConvNet: CNNs for MATLAB
Other
1.4k stars 753 forks source link

vl_nnconv error when training CIFAR with GPU #311

Open tinalegre opened 8 years ago

tinalegre commented 8 years ago

Hi everybody,

I compiled v16.0 of matconvnet with GPU support. I was trying to train the CIFAR example using GPU, i.e.:

opts.train.gpus=[1];

Right away on the very first iteration, following error appears: Cifar Error

Has anybody experienced the same problem? I'm using CUDA v7.5 + Matlab 2014a. What it's strange, is that when I run the example on cnn_mnist with GPU modus, no errors occur and the training is done without problems: Mnist

Any ideas are welcomed :-)

iiwindii commented 8 years ago

your DriverVersion (cuda) is 7.5 while your ToolkitVersion (matlab 2014 ) is 5.5.... The two should had better match. The ideal configuration is "R2014a+cuda 5.5", "R2014b+cuda 6.0", "R2015a+cuda 6.5"...... Your GPU is K40, so you can try "R2015a+cuda 6.5". hope this can help you

tinalegre commented 8 years ago

@fengyunxiaozi thanks! do you have a link about those ideal configurations? I recently installed Matlab2015b, from this link to Matlab, I could find that for Matlab2015b is Cuda v 7.0, however I couldn't find information for previous Matlab versions. On the other hand, does it means that CUDA v7.5 shall be avoided with Matlab 2015b?

iiwindii commented 8 years ago

http://www.vlfeat.org/matconvnet/install/ here you can find the officially suggested configuration. I think you had better follow these recommended configurations