# DepthWise Convolution. Apply two different vector weights to two different input data with one conv

Hi,

Before start explaining the problem I would like to clarify that I am simplifying the whole problem to make easy the question. What I am trying to do is: I have 2 vectors of 1x3x5 (batch, feat, lenght) as input data. I need to apply a different filter to each vector of size (output_channels, in_channels, lenght) in our case (12,3,5). I would like to remark that this filter is different for each of the samples. In order to apply one of this filter to one of the vectors I am using the torch.nn.functional.conv1d function as you can see below:

data = torch.randn(1,3,5)
weight = torch.randn(12,3,5)
out = torch.nn.functional.conv1d(data,weight)
data.size()
torch.Size([1, 3, 5])
weight.size()
torch.Size([12, 3, 5])
out.size()
torch.Size([1, 12, 1])

The most straighforward to apply this function for each vector taking intou account that the weight is different for each one is using a for loop, but it is inneficient. I was thinking to use the depth wise convolution using the groups parameter, but I am not sure if I have understood properly the depth wise convolution properly. Below you can see a snippet.

data_1= torch.randn(1,3,5)
data_2= torch.randn(1,3,5)
data = torch.cat((data_1, data_2), dim=1)
data.size()
torch.Size([1, 6, 5])
weight_1 = torch.randn(12,3,5)
weight_2 = torch.randn(12,3,5)
weight = torch.cat((weight_1, weight_2), dim=1)
weight.size()
torch.Size([12, 6, 5])
out = torch.nn.functional.conv1d(data,weight,groups=2)
Traceback (most recent call last):
File “”, line 1, in
RuntimeError: Given groups=2, weight of size [12, 6, 5], expected input[1, 6, 5] to have 12 channels, but got 6 channels instead

Basically I concatenate the output vectors in the dimension of the feature dimension and the weight vectors in the dimension of the input feature. I expected that using groups = 2 to have the same behavior as applying the convolution individually. But it is not working Can you tell me what I’m missing? Is there anyway to implement my usecase using the current convolution implementation?