Custom kernel for Convolution

vision
1

I want to perform convolutions with custom kernels, but in a manner different than traditional convolutions. While convolution in pytorch occurs in general by each kernel slid and multiplied with each channel and then the sum passed over, I need to perform convolutions where input and output channels are same and each kernel (same as the number of channels) is multiplied with only a single channel from the input. And concatenated with the rest of the outputs from the other kernels. It looks like this now:

in_channels=out_channels
filters = torch.zeros((batch_size,in_channels,out_channels,f_size,f_size)
for batch in range(batch_size):
     for chan in range(in_channels):
            filter_use=filters[batch,chan].unsqueeze(0)
            out_conv=F.conv2d(x[batch, chan].unsqueeze(0).unsqueeze(0), filter_use)
            
            ### Perform more operations
### Concatenate the results

While this gets the work done, it takes a huge amount of time as you can guess from the nested loop structure. What would be a faster alternative to this?

2

Boosting this ^^. Can anyone help me please :slight_smile: Thanks!

3

This sounds like a grouped convolution with the number of groups == number of input channels Conv2d — PyTorch 1.9.0 documentation, but I’m not sure because the current code example doesn’t run. Could you post a runnable code snippet to help confirm? The indexing of the filters by batch also seems very suspicious, as it implies that different elements of the batch are treated differently…

2 Likes
4

Thanks! It is indeed grouped convolutions. One more question–Grouped convolutions “assigns” kernels to channels in the order of their indices, right?

5

This sounds like it should be the case. As always, you can test this theory with synthetic data/weights to be sure.