python - Theano: mixing CPU and GPU? -
i built neural network needs use cholesky decomposition , solve triangular systems part of computation. means need compute gradients of whole computation, of course.
when try compile code error "no cula available". unfortunately, can't download cula website.
i wondering if it's possible mix cpu , gpu theano. matrices need use cholesky , solve on small (100x100) part on cpu. so, though, i'd need transfer matrices cpu right after they've been computed gpu , send result gpu. possible transparently enough?
this transfer transparent. no need special. recommand profile theano graph, sure: http://deeplearning.net/software/theano/tutorial/profiling.html#tut-profiling
for cula, there pr give solve op based on cusolver provided nvidia. not needed anymore. try pr: https://github.com/theano/theano/pull/4917
Comments
Post a Comment