I think you misunderstood something. The current/old gpu back-end only support float32. It don't support any int. Not even int32. There is a few work around internally, but if you need to work on integer use the new gpu back-end:
http://deeplearning.net/software/theano/tutorial/using_gpu.html#gpuarray-backend Fred On Tue, Oct 25, 2016 at 5:37 AM, Ghaelen88 <[email protected]> wrote: > > > <https://lh3.googleusercontent.com/-ukMswfeSQkI/WA8nWuqfe6I/AAAAAAAAAYE/A4a99ROSrw0twSELiZzxiuPbjv5cz2pbwCLcB/s1600/Screen%2BShot%2B2016-10-25%2Bat%2B10.34.34.png> > > > <https://lh3.googleusercontent.com/-gVilakFH7o8/WA8m_S0GwFI/AAAAAAAAAYA/xX3f-7ubL5k_chOMtpLKmtY-YJyN003RQCLcB/s1600/Screen%2BShot%2B2016-10-25%2Bat%2B10.27.37.png> > > > > > <https://lh3.googleusercontent.com/-MAhiwbUKS38/WA8mDT08o4I/AAAAAAAAAX4/_tOGQI3aTOQdR17iVlzSgyVwidhk1dr-wCLcB/s1600/Screen%2BShot%2B2016-10-25%2Bat%2B10.27.37.png> > Hi everyone, I'm trying to run some code entirely on the GPU, but when I > print the symbolic graph it seems some operations are still happening on > the CPU are they are using int64 values (which I guess need the CPU as GPU > supports just int32. In particular, it seems that the sub-tensor operation > that goes then into the scan. > > So I have two questions (related to the two images above): > > 1) How can I make sure I use the sub-tensor operation with int32 and only > shared variables? > 2) If the scan node says for{CPU, scan_fn} does that mean I'm still using > mainly the CPU? > > Thanks so much! > > > > -- > > --- > 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.
