A decoder for Semantic Segmentation

[cschloer]

Hi,

I am interested in using your architecture for a semantic segmentation problem. I am therefore using the segmentation_models.pytorch library, which luckily implements timm and therefore your architecture as the encoder.

However, all of the decoders supported by segmentation_models.pytorch use normalization. Should I just replace all instances of Conv2D followed by BatchNorm2D with a ScaledStdConv2D, or do you have a better suggestion? (Should I also then put the ReLU before the ScaledStdConv2D, as you seem to do?)

Thank you in advance.