-
CUDA issues
If MagmaDNN is building in CPU-only mode, but you're expecting it to compile for the GPU, then make sure the CUDA toolkit is in your path. Identify where your cuda installation is located and run
export PATH="${PATH}:/path/to/cuda/bin"
in your shell.'Cannot find header "cudnn.h"' or similar errors. Make sure you have CuDNN installed. See this site for more info on installing CuDNN.
'Cannot find header "magma.h"' or similar errors. Magma must be installed in order to use MagmaDNN. See this site for more info on installing Magma.
-
Runtime Errors
Stack smashing detected -- Currently there is a known memory bug in MagmaDNN revolving around the
::magmadnn::model::NeuralNetwork<T>::fit
function. If you pass stack allocated tensor objects as thex
andy
(i.e.Tensor<float> x (...);
), then somewhere in the network training, the stack is accessed out-of-bounds. MagmaDNN 1.1 or 1.2 aims to have an overhauled memory system, which handles this issue. As a workaround, please pass heap allocated tensor objects to the fit function (i.e.Tensor<float> *x = new Tensor<float> (...);
).