Circular padding in convolutional neural networks

[VolodyaCO]

Description of the model to be implemented

In many areas such as physics, it is convenient to have convolutional layers with periodic boundary conditions (e.g. see netket)

Therefore, it would be nice to add a "CIRCULAR" padding option to convolutional layers, just as they do in neural-tangents.

Dataset the model could be trained on

1D or 2D data. Maybe MNIST images.

Specific points to consider

None in particular. Just as an example, suppose that one has the 1D data [1,2,3,4,5] and one has filters of size 3, and a stride of 3. The idea is then that two filter operations are carried out. The first one will use [1,2,3], and the second one will use [4,5,1].

Reference implementations in other frameworks

neural-tangents has replaced stax's GeneralConv by a Conv layer, which has this padding option, and further does not require to provide directly the XLA's dimension_numbers.

[marcvanzee]

I think it would be quite nice to add this, since it doesn't seem to complicate the API much (no additional parameters etc). @levskaya what do you think of this proposal? I recall you were involved in a discussion around this before, and I'm curious whether you think it makes sense to add this.

[jheek]

It would be even nicer if the jax conv op would support this out of the box. They already have 'same' and 'valid'.

[sgrigory]

If this is still relevant, I'd be happy to raise a PR, reusing the code from #903 (comment) and adding some tests

[VolodyaCO]

I'd love it if you do that Grisha!

…

On Wed, 3 Nov 2021 at 17:55, Grisha Oryol ***@***.***> wrote: If this is still relevant, I'd be happy to raise a PR, reusing the code from #903 (comment) <#903 (comment)> and adding some tests — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#971 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AHQJA72FDQBHK5DPXFZCKLTUKG4XLANCNFSM4W4YKQGQ> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.

-- Best wishes, Vladimir Vargas-Calderón <https://www.researchgate.net/profile/Vladimir_Vargas-Calderon> PhD Physics Student @ Universidad Nacional de Colombia

-- *Aviso legal:* El contenido de este mensaje y los archivos adjuntos son confidenciales y de uso exclusivo de la Universidad Nacional de Colombia. Se encuentran dirigidos sólo para el uso del destinatario al cual van enviados. La reproducción, lectura y/o copia se encuentran prohibidas a cualquier persona diferente a este y puede ser ilegal. Si usted lo ha recibido por error, infórmenos y elimínelo de su correo. Los Datos Personales serán tratados conforme a la Ley 1581 de 2012 y a nuestra Política de Datos Personales que podrá consultar en la página web  www.unal.edu.co <http://www.unal.edu.co/>.* *Las opiniones, informaciones, conclusiones y cualquier otro tipo de dato contenido en este correo electrónico, no relacionados con la actividad de la Universidad Nacional de Colombia, se entenderá como personales y de ninguna manera son avaladas por la Universidad.