The OCR approach is rephrased as Segmentation Transformer: https://arxiv.org/abs/1909.11065. This is an official implementation of semantic segmentation for HRNet. https://arxiv.org/abs/1908.07919
I was wondering if the authors experimented with LayerNorm instead of BatchNorm and if they noticed any differences? From what I've read is that BatchNorm is heavily dependent on batch size. My max batch size while using this architecture is 32 while using 2 A100 GPUs and this is why I had the idea of replacing BatchNorm with LayerNorm.
I was wondering if the authors experimented with LayerNorm instead of BatchNorm and if they noticed any differences? From what I've read is that BatchNorm is heavily dependent on batch size. My max batch size while using this architecture is 32 while using 2 A100 GPUs and this is why I had the idea of replacing BatchNorm with LayerNorm.
Did anyone else try it?