Thanks guys, I've spent a bunch of time playing convnet sudoku and I'm 
converging towards agreement that there's no simple way to do it with 
tricks or reshapes. I think you might be able to do some sort of 
factorization where you do an elementwise multiply that spans the batch 
dimension but since I'm just exploring I think I'll go with Michael's 
solution. Thanks again!

-Andy

On Tuesday, December 13, 2016 at 10:24:26 PM UTC, Pascal Lamblin wrote:
>
> Hi, 
>
> Unfortunately, I don't think there is an easy way yet. 
>
> If the batch size is small enough, you can try to define a "diagonal" 
> set of 3D filters and use 3D convolutions. 
>
> Otherwise, if we have "grouped" convolutions, that might help, but I 
> think none of the current convolution back-ends implement that. 
>
> On Tue, Dec 13, 2016, Andrew Brock wrote: 
> > Hi guys, 
> > 
> > I'm doing some work where I'd like to have a data-dependent 
> convolutional 
> > layer, where a different filterbank is used on each element of a given 
> > batch. This is easy when doing things with a single batch, but for 
> > minibatches I'd end up with a 5D  Batch x Num_filters x Channels x H x W 
> > tensor, and I can't think of an easy way to apply this on a 4D Batch x 
> > Channel x H x W tensor. Is there some trick with backward passes, 
> > dimshuffles and 3D reshapes, or something else that anyone knows of that 
> > would give me an efficient way to do this? 
> > 
> > Thanks, 
> > 
> > Andy 
> > 
> > -- 
> > 
> > --- 
> > 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] <javascript:>. 
> > 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 [email protected].
For more options, visit https://groups.google.com/d/optout.

Reply via email to