1d convolution
There api don't support our extra dimensions(batch, nkern, stacksize)
so we can't deprecate our current api.
But we could have it to be more compatible if people have time to do this.
Fred
- Hide quoted text -
On Tue, Jul 26, 2011 at 9:59 AM, James Bergstra
wrote:
> thoughts on borrowing aspects of scikits ndimage API?
>
>
> ---------- Forwarded message ----------
> From: James Bergstra
> Date: Tue, Jul 26, 2011 at 9:59 AM
> Subject: Re: [theano-users] 1d convolution op
> To: theano-users@googlegroups.com
>
>
> There was some discussion of this a few months ago on the mailing
> list. I think the upshot was that the 2d convolution can be used as-is
> to do 1d convolution with reasonable performance (provided the filter
> is small enough).
>
> It would be nice to mirror something like the scikits ndimage
> interface to provide 1d, 2d, and Nd convolution and correlation
> functions. These are all wrappers around some of the same underlying
> computational techniques, it's just a matter of putting a nicer API on
> top of them.
>
> On Tue, Jul 26, 2011 at 8:57 AM, jzhang wrote:
>> Does theano have a built in 1d convolution op like the 2d convolution
>> op? I'm testing a CNN on a time series dataset.
>
so we can't deprecate our current api.
But we could have it to be more compatible if people have time to do this.
Fred
- Hide quoted text -
On Tue, Jul 26, 2011 at 9:59 AM, James Bergstra
wrote:
> thoughts on borrowing aspects of scikits ndimage API?
>
>
> ---------- Forwarded message ----------
> From: James Bergstra
> Date: Tue, Jul 26, 2011 at 9:59 AM
> Subject: Re: [theano-users] 1d convolution op
> To: theano-users@googlegroups.com
>
>
> There was some discussion of this a few months ago on the mailing
> list. I think the upshot was that the 2d convolution can be used as-is
> to do 1d convolution with reasonable performance (provided the filter
> is small enough).
>
> It would be nice to mirror something like the scikits ndimage
> interface to provide 1d, 2d, and Nd convolution and correlation
> functions. These are all wrappers around some of the same underlying
> computational techniques, it's just a matter of putting a nicer API on
> top of them.
>
> On Tue, Jul 26, 2011 at 8:57 AM, jzhang wrote:
>> Does theano have a built in 1d convolution op like the 2d convolution
>> op? I'm testing a CNN on a time series dataset.
>
Leave a comment
The thread with impact on GPU performance:
https://groups.google.com/group/theano-users/browse_thread/thread/2491d99ae50374f1/dbb95fd3fd3cad4a?hh1. en_USlnkgstq=1d+convolution* dbb95fd3fd3cad4a
https://groups.google.com/group/theano-users/browse_thread/thread/2491d99ae50374f1/dbb95fd3fd3cad4a?hh1. en_USlnkgstq=1d+convolution* dbb95fd3fd3cad4a