Dilated ResNet in SFNet?

Hi, I find that the SFNet w/ ResNet-18 has adopt dilated convolutions in C4 and C5:

Line 202 in 773c7b7

    
           return AlignNetResNet(num_classes, trunk='resnet-18-deep', criterion=criterion, variant='D', skip='m1', fpn_dsn=True)

However, it hasn't been mentioned in the paper or the other implementation torchcv.
I'm confused about it.

@wondervictor Hi! Thanks for your remind. Note that this is re-implementation of our SFnet which achieve better reasuts than paper. You can turn the dilation off. I remember it only drop a little after longer training.

Hi @lxtGH, I found another setting was strange to me.
see:

SFSegNets/network/sfnet_resnet.py

Line 53 in ce97d2e

conv3x3_bn_relu(fpn_dim, fpn_dim, 1),

and

SFSegNets/network/sfnet_resnet.py

Line 79 in ce97d2e

conv3x3_bn_relu(len(fpn_inplanes) * fpn_dim, fpn_dim, 1),

The normalization layers in fpn_out and conv_last are nn.BatchNorm2d but not SyncBatchNorm due to unspecific parameter normal_layer for function conv3x3_bn_relu.
I'm concerned about whether it is a special setting or a bug.

It may be a bug but it seems that it dose not effect the performance.

Time to close this issue. If you have more question. please re-open it.