CNN: array dimensions from convolution make no sense to me

vision
1

Hi. I have a tensor which is made up of 8 matrices, and each matrix is 20x20 numbers. I am trying to do a convolution on this matrix but I don’t understand the output dimensions. Here is what I have so far (the x is my tensor):

class CNN(nn.Module):

def __init__(self, n_filters, filter_sizes):

    super(CNN, self).__init__()

    self.conv0 = nn.Conv2d(in_channels=1, out_channels=n_filters[0], kernel_size=filter_sizes[0])

def forward(self, x, lengths=None):
    x = x.unsqueeze(1)

    map0_conv = F.relu(self.conv0(x))
    map0_avg = torch.mean(map0_conv, dim=1)

My filter size is 2x2, and the stride is 1. Therefore, for each matrix I need 19*19 = 361 filters. Thus, starting with an 8x20x20 tensor, I end up with the dimensions of self.conv0(x) being 8x361x19x19.

I don’t know what this result means. It’s causing me problems in the next step, the torch.mean. What I’m trying to do, is to take a 8x20x20 tensor, apply a convolution to it and then get the average of each convoluted matrix to end up with one vector containing 8 entries. Clearly I’m doing it wrong but I don’t know how to fix it.

2

Hi,

So, when you are doing x.unsqueeze(1), you are basically passing an input size as (8,1,20,20). Now, the output dimensions of your conv operation depends on the number of out_channels you pass in.
so, let’s say you pass in out_channels=1, then you end up with an output size of (8,1,19,19). [19 because you have a kernel of size 2x2 and stride as 1]. In short, the output size (8, x, 19, 19) where x is a variable depending on the number of out_channels you pass.

Finally, for the mean to be a vector of 8 entries, all you need to do is calculate the average across axis=1,2,3.

map0_avg = torch.mean(map0_conv, dim=[1,2,3])

the results would be a vector of size 8.