If it return Theano, then Theano is able to compile again cuDNN as Doug told. dnn.library_path is a default that we try to use. But even if cuDNN isn't there and is well installed, Theano will find it and use it. We just don't try to put in dnn.library_path the exact path of cuDNN. But this allow user to tell Theano where it is if they need to.
Fred On Tue, Jul 12, 2016 at 11:30 AM, Doug <[email protected]> wrote: > IIRC theano checks for CuDNN by attempting to actually compile against it, > so if it can then it's finding the CuDNN files somewhere. > > > On Tuesday, July 12, 2016 at 11:23:02 AM UTC-4, [email protected] > wrote: >> >> My theano.sandbox.cuda.dnn_available() says "True". >> >> However, I noticed that theano.config.dnn.library_path == >> '/usr/local/cuda/lib64', which is a directory that has no cuDNN files on my >> system. >> >> If Theano is looking for cuDNN in the wrong place, how can cuDNN be >> available to it? >> >> Is there any way to know for cure that Theano is using cuDNN in its >> convolution operations? The speed that I'm seeing seems to be about a >> quarter of what cuDNN should be capable of. >> > -- > > --- > 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 [email protected]. > For more options, visit https://groups.google.com/d/optout. > -- --- 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 [email protected]. For more options, visit https://groups.google.com/d/optout.
