It may be possible that ultra_fast_sigmoid has no implementation on the GPU, or that the corresponding optimization is not applied for some reason.
On Tue, Dec 06, 2016, Bogdan Budescu wrote: > Could it be that when specifying > optimizer_including=local_ultra_fast_sigmoid in theano's config flags, the > gpu optimization is disabled? > > On Tuesday, December 6, 2016 at 5:40:23 PM UTC+2, Bogdan Budescu wrote: > > > > When trying to compile a (rather large) neural net with device=cuda, at > > some point I get the following error: > > > > ImportError: ('libamdlibm.so: cannot open shared object file: No such file > > or directory', '[Elemwise{pow,no_inplace}(<TensorType(float32, scalar)>, > > <TensorType(float32, scalar)>)]') > > > > Now, I don't mind the error itself, as it's probably caused by not running > > ldconfig, but this suggests that the optimizer might want to run the op on > > cpu instead of gpu, and I assume that this might be the reason for which my > > net's training runs so slow (I assume that this also implies some redundant > > copies between cpu and gpu memory). I also observe that the training > > process takes 100% of the processor (or, rather, of a single core, as > > python is not multi-threaded). > > > > How can I tell whether there are any ops running on cpu after a successful > > compilation (I already set assert_no_cpu_op='raise'), and how can I force > > the ops to be executed on gpu instead? > > > > > > > > -- > > --- > You received this message because you are subscribed to the Google Groups > "theano-users" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to theano-users+unsubscr...@googlegroups.com. > For more options, visit https://groups.google.com/d/optout. -- Pascal -- --- You received this message because you are subscribed to the Google Groups "theano-users" group. To unsubscribe from this group and stop receiving emails from it, send an email to theano-users+unsubscr...@googlegroups.com. For more options, visit https://groups.google.com/d/optout.