Hi nouiz,

If you can make it work, I for one would really appreciate if it was 
possible to pool an image where the output of the pooling operation has 
more than 512 columns. I need that since in my network I pool only 
vertically and not horizontally. I instead use dilated convolution layers 
where I dilate in the x-direction; this is in order not to loose any 
spatial information horizontally which is important in my case. For now I 
will have to crop the image to 512 columns but I would really prefer not to.

Regards

Kristofer

On Friday, May 13, 2011 at 5:31:43 PM UTC+2, nouiz wrote:
>
> Hi,
>
> It is an NVIDIA gpu card limitation. But we can work around by reusing
> thread for more then one columns.
>
> I updated the gpu function to work around this limitation. So the
> gradient of the downsample work when the input of the downsample have
> more then 512/1024 columns. There is still the limitation that the
> output of the downsample must have less then 512 colomns.
>
> If you need it, I can try to make it. Otherwise, it will wait.
>
> You must use and update the version on mercurial to have the new 
> functionnality.
>
> HTH,
>
> Fred
>
> On Fri, May 13, 2011 at 10:34 AM, Sander Dieleman
> <[email protected] <javascript:>> wrote:
> > Oh, that's great! I figured this was just an inherent limitation of
> > the CUDA platform, but if it can be fixed that's even better :)
> >
> > On May 13, 3:20 pm, Frédéric Bastien <[email protected]> wrote:
> >> Hi,
> >>
> >> I know how to fix the problem in the code. That way, you won't need
> >> workaround and it will work in all case.
> >>
> >> I will try to put a better error message. The error that you see is
> >> the error from nvidia that we forwarded.
> >>
> >> thanks for the report.
> >>
> >> Frédéric Bastien
> >
>
>

-- 

--- 
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.

Reply via email to