syncedmem.cpp:71] Check failed: error == cudaSuccess (2 vs. 0) out of memory

[Shrabani Ghosh]

I0713 13:16:30.953454  2550 net.cpp:242] This network produces output recon

I0713 13:16:30.953469  2550 net.cpp:255] Network initialization done.

I0713 13:16:30.964470  2550 upgrade_proto.cpp:79] Attempting to upgrade batch norm layers using deprecated params: /home/sghos003/Desktop/caffe-master/build/install/python/SRCNN_cmr_upsample_x5.caffemodel

I0713 13:16:30.964491  2550 upgrade_proto.cpp:82] Successfully upgraded batch norm layers using deprecated params.

I0713 13:16:30.964592  2550 net.cpp:744] Ignoring source layer loss

F0713 13:16:31.361512  2550 syncedmem.cpp:71] Check failed: error == cudaSuccess (2 vs. 0)  out of memory

*** Check failure stack trace: ***

Aborted (core dumped)

is there any way to free the space to fix this problem? 

[Przemek D]

This error refers to your GPU memory, apparently your card doesn't have sufficient RAM to contain the network. The easiest solution would be to decrease the batch size. If you're already at size 1, you could try to use smaller images. Also, remember that some solvers need more memory than others (e.g. Adam needs more than SGD). Other than that the suggestions get more difficult... you could either remove some layers from the network or, well, buy a bigger GPU.