convolution layer with cardinality

[OValery16]

Dear author,

Thank for this impressive piece of work. How to implement convolution layer with cardinality ?

in universal.py we have:

def melt(self):
        if self.conv.groups == 1:
            groups = 1
        elif self.conv.groups == self.conv.out_channels:
            groups = int((self.out_mask != 0).sum())
        else:
            assert False

        replacer = nn.Conv2d(
            in_channels = int((self.in_mask != 0).sum()),
            out_channels = int((self.out_mask != 0).sum()),
            kernel_size = self.conv.kernel_size,
            stride = self.conv.stride,
            padding = self.conv.padding,
            dilation = self.conv.dilation,
            groups = groups,
            bias = (self.conv.bias is not None)
        ).to(self.conv.weight.device)

        with torch.no_grad():
            if self.conv.groups == 1:
                replacer.weight.set_(self.conv.weight[self.out_mask != 0][:, self.in_mask != 0])
            else:
                replacer.weight.set_(self.conv.weight[self.out_mask != 0])
            if self.conv.bias is not None:
                replacer.bias.set_(self.conv.bias[self.out_mask != 0])
        return replacer

if the convolution layer have a cardinality (like in several modern model), we get assert False. How to implement cardinality for convolution layers ?

[youzhonghui]

Removing a single filter in the group convolution in PyTorch will cause misalignment, but it's possible to remove the entire group by using the Group Pruning method proposed in our paper.

[OValery16]

Thanks a lot for your quick reply.

I understand that using Group Pruning method allows us to synchronize the number of channel that are pruned within a group in order to avoid misalignment. However, I am still confused how to make sure the final input number of channel in a given Group Pruning is dividable by the number of group (in the convolution)

Should I modify g.minimal_filter such as g.minimal_filter=int(g.minimal_filter//conv_groups) ?

[OValery16]

In fact, I am a bit confused how to remove an entire group.

[youzhonghui]

If you remove a whole group of filters at a time, the number can be kept divisible. The troublesome thing is to maintain the number of channels of the input feature map, since group convolution also divides input. For example, given 6 filters that are divided into 3 groups, and after removing the first group of convolutions, there are 4 filters left and the cardinality should change to 2. The input channel may still be 6, and two corresponding feature maps need to be discarded. So the code to prune a network with cardinality may be very different from what we provided.