Thanks for your consistent updates on this repo. I notice that in the update for occ task, there is a SyncBNHook implemented for models with STBase as their backbones. It is described that SyncBNHook can prevent loss divergency. Is there a connection between divergent loss and normal BatchNorm2D? I wonder if you can share with us your opinions on this conversion.
@woxihuanjiangguo Just empirical setting…… BN is used in both BEVEncoder and Head. The batch size in training STBase is too small to obtain stable statistics.
Thanks for your consistent updates on this repo. I notice that in the update for occ task, there is a SyncBNHook implemented for models with STBase as their backbones. It is described that SyncBNHook can prevent loss divergency. Is there a connection between divergent loss and normal BatchNorm2D? I wonder if you can share with us your opinions on this conversion.