Number of output channels of conv_offset_mask

[bluesky314]

The number of channels for offset and modulation usually has 27 channel output as kernal size is 3. But according to my understanding of paper, for each channel we a 3 channel predictions i.e shift in x,y and the modulation factor. So the output channels should be 3 times the input however that is not what is coded in:

self.conv_offset_mask = nn.Conv2d(self.in_channels,
                                          self.deformable_groups * 3 * self.kernel_size[0] * self.kernel_size[1],
                                          kernel_size=self.kernel_size,
                                          stride=(self.stride, self.stride),
                                          padding=(self.padding, self.padding),
                                          bias=True)

The paper says

The output is of 3K channels, where the first 2K channels correspond to the learned offsets and the remaining K channels are further fed to a sigmoid layer to obtain the modulation scalars

Can someone clarify what I am missing and because this layer outputs 27 channels no matter the input dimension so how is this used for shift x,y and modulation?