search

vlfeat / matconvnet

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

running cnn_cifar on GPU #142

Closed charlesboes closed 9 years ago

charlesboes commented 9 years ago

Hi, I am having some trouble running can_cifar on a GPU unit. the compilation of Matconvnet with GPU enabling seemed to have worked, but when I am trying to run can_cifar I get this message of error:

Error using vl_nnconv nnconv_forward_blas<>: : gemm<>: [cublas error] [cublas:cublasSgemm (CUBLAS_STATUS_EXECUTION_FAILED)]

I get the same type of error running can_mnist.

I am using Matlab R2013a CUDA-7.0 GPU: 'GeForce GTX 980'

Best Regards

charlesboes commented 9 years ago

Actually it seems like it's a cublas issue which is an cudaErrorInvalidTexture. You can get rid of this error using cudNN implementation. One may also want to comment l.350-367 of vl_nnconv.cu to avoid using cublas for fully connected layers.

RomainBUAA commented 7 years ago

i use cudnn ,but this error doesn't disappear