Dimensions

[albertoandreottiATgmail]

Really simple change to allow the dimensions of the input to be different. In this way, one of the dimensions can be seen as time, so you can build a TDNN(time delay neural network) to work with speech signals.

[rasmusbergpalm]

please calculate ones(xscale, yscale) / (xscale*yscale) in cnnsetup, put it in the net somewhere (net._meanFilter or similar), use that and remove the comment to replace with var

[rasmusbergpalm]

Funky.. #76 implements non-square kernels. Did you guys coordinate?

Anyway, same comments apply.
make scale a vector, don't be backwards compatible, fix examples to use new style and add assertions in cnnsetup to verify correct scale (i.e. positive integer vector) in order to fail fast.
Add a test that using non-square and square scale works as intended. (i.e. the outputmaps are the expected size, for a known input the computed output is correct, etc.)

Sorry about the extra feedback btw. I was really busy last time I reviewed.

[rasmusbergpalm]

Any updates on this?

[albertoandreottiATgmail]

Hi guys,

I believe this is already working in a branch of my clone of the repo. Let
me check :-).

On 11 May 2014 10:01, Rasmus Berg Palm [email protected] wrote:

Any updates on this?

—
Reply to this email directly or view it on GitHubhttps://github.com//pull/75#issuecomment-42769948
.

José Pablo Alberto Andreotti.
Tel: 54 351 4730292
Móvil: 54351156526363.
MSN: [email protected]
Skype: andreottialberto

[rasmusbergpalm]

What's the status?