DCNV2 MODULATOR WEIGHT

[zhenyuzhouiu]

PyTorch-Deformable-Convolution-v2/dcn.py

Line 54 in 4d958eb

modulator = 2. * torch.sigmoid(self.modulator_conv(x))

I have a question about the modulator weight. Why do you use 2 to multiply the result of the sigmoid() instead of 1? Shouldn't this modulator range from 0 to 1?

[developer0hye]

@zhenyuzhouiu
Good question

I followed the original implementation, but I don't know why it is designed that way.

I also uploaded this issue to the official implementation project, but I couldn't get the answer.

msracver/Deformable-ConvNets#280

I got improved results by multiplying 2 by a modulator.