00:33 <ShubhamAgrawal[m> <rcurtin[m]> "Shubham Agrawal: you didn't..." <- I think it was about weight

00:33 <ShubhamAgrawal[m> You can see that we are using weight.slice(outMap) which will just iterate from 0 to maps-1 only

03:04 <rcurtin[m]> Right, but there is only a 2D weight for each map. The same 2D weight is applied to all the higher dimensions of a particular input. But, it seems that this might not be the behavior you expected? I don't have a problem with holding one weight map for each higher input dimension, if you want to change the code

04:27 <ShubhamAgrawal[m> But you initialized weight dimensions as

04:27 <ShubhamAgrawal[m> `MakeAlias(weight, weightPtr, kernelWidth, kernelHeight, maps * inMaps);`

04:28 <ShubhamAgrawal[m> I think we are doing something wrong here

04:38 <ShubhamAgrawal[m> <rcurtin[m]> "Right, but there is only a 2D..." <- Actually I am not talking about higher dimension input here

04:38 <ShubhamAgrawal[m> See this

04:38 <ShubhamAgrawal[m> https://datascience.stackexchange.com/questions/67318/convolution-layer-dimensions-in-deeper-layers

13:13 <rcurtin[m]> how is that response different than what I suggested above? I am not following what you mean; can you please clarify the precise problem and what the fix would be?

13:16 <ShubhamAgrawal[m> > <@shubhamag:matrix.org> I think it was about weight... (full message at https://libera.ems.host/_matrix/media/r0/download/libera.chat/8b8c84fe7123028f4a8692b7d45e0a61d2992e98)

13:19 <ShubhamAgrawal[m> Modify `weight.slice(outMap)`

13:19 <ShubhamAgrawal[m> to `weight.slice(outMap * inMaps + inMap)`

13:20 <ShubhamAgrawal[m> > <@shubhamag:matrix.org> Modify `weight.slice(outMap)`

13:20 <ShubhamAgrawal[m> > to `weight.slice(outMap * inMaps + inMap)`

13:20 <ShubhamAgrawal[m> Something on these lines

13:25 <rcurtin[m]> yes, I see what you mean now; do you want to open a PR?

13:26 <ShubhamAgrawal[m> I actually want to implement groups in convolution layer

13:26 <ShubhamAgrawal[m> What do I need to change in this case?

13:26 <rcurtin[m]> I'm not familiar with grouped convolutions; you should implement it as an entirely new layer, though

13:27 <ShubhamAgrawal[m> rcurtin[m]: why

13:27 <ShubhamAgrawal[m> > <@ryan:ratml.org> I'm not familiar with grouped convolutions; you should implement it as an entirely new layer, though

13:27 <ShubhamAgrawal[m> * why?

13:28 <rcurtin[m]> well, like I said, I am not familiar with grouped convolutions; but assuming it is a different operation than the regular convolution layer, it is better to keep the `Convolution` class as simple as possible

13:29 <ShubhamAgrawal[m> rcurtin[m]: I think it will require to just adjust loops

13:32 <rcurtin[m]> you're welcome to make the change and open a PR and see what it looks like, but be aware I may request it to be an entirely separate layer. it's important to keep each layer simple: as one related example, it's better to have a separate layer for convolution and transposed convolution, instead of adding some flag to the convolution class that toggles transposed convolution mode. This is because of the huge amount of increased code complexity

13:32 <rcurtin[m]> of handling that flag, and also that it makes it harder for a new person to come along and quickly understand that code---if it is two separate classes, they can understand each class as performing one operation (convolution or transposed convolution), but if it's one, they have to understand both

13:33 <ShubhamAgrawal[m> > <@shubhamag:matrix.org> Modify `weight.slice(outMap)`

13:33 <ShubhamAgrawal[m> > to `weight.slice(outMap * inMaps + inMap)`

13:33 <ShubhamAgrawal[m> And is this right?

13:33 <ShubhamAgrawal[m> Or do I need to write something else?

13:33 <rcurtin[m]> yes, I think you are correct about the change to `weight` there