How does the code in the psconv.py file achieve the different dilation rates written in the paper?

[xlh1135355585]

How does the code in the psconv.py file achieve the differentdilation rates written in the paper?

[doitslow]

How does the code in the psconv.py file achieve the differentdilation rates written in the paper?

I think the different dilation rates are done through the three different convs. However, I find it very difficult to understand how the three conv results add up together and result in the 1,2,1,4 cyclic dilation rates. I can see how half of the weights are been set off in the conv with no dilation. But what is the reordering of channels in shift conv for? @d-li14 Possible to take some time to do a little bit elaboration? It would be very appreciated.

[urbaneman]

How does the code in the psconv.py file achieve the differentdilation rates written in the paper?

I think the different dilation rates are done through the three different convs. However, I find it very difficult to understand how the three conv results add up together and result in the 1,2,1,4 cyclic dilation rates. I can see how half of the weights are been set off in the conv with no dilation. But what is the reordering of channels in shift conv for? @d-li14 Possible to take some time to do a little bit elaboration? It would be very appreciated.

I have the same question. Look forward to the author's answer. @d-li14

https://github.com/d-li14/PSConv/blob/9afde67f43025bd5ae59e1ff1cc5a04c0744a5bd/mmdet/models/utils/psconv.py#L21

[Margrate]

but the function didn't use the line: self.mask[i * _out_channels: (i + 1) * _out_channels, i * _in_channels: (i + 1) * _in_channels, : , :] = 

def forward(self, x): x1, x2 = x.chunk(2, dim=1) x_shift = self.gwconv_shift(torch.cat((x2, x1), dim=1)) return self.gwconv(x) + self.conv(x) + x_shift

just add three different conv. It‘s not the same with paper

[XinzheGeng]

In my opinion, x_shift = self.gwconv_shift(torch.cat((x2, x1), dim=1)) can achieve conv with dilation 4 (green part in paper section B fig. 4), self.gwconv(x) can achieve conv with dilation 2 (red part in paper section B fig. 4), self.conv(x) can achieve conv with dilation 1 (blue and yellow parts in paper section B fig. 4). where the param 'dilation' for class PSConv2d have to be 2 instead of 1. It's a may simple implementation for this method.