Dear All,
Sometimes I got "Segmentation fault (core dump)" error message while training (with Cuda 7.0). If I reboot the machine everything works fine.
I don't know if this is caused by some unfreed GPU memory? Or garbagecollect does free GPU memory as well?
Is there a way to call cudaDeviceReset() via Torch7 / cutorch?
Any hints are highly appreciated!
Best regards,
Balint Toth