Is it ok to directly replace the BN with GN(group normalization) in a BN-pretrained model(I don’t have time to train a GN-based model on ImageNet), and finetune the model for object detection? e.g. finetune on COCO, will there be a big difference between the performance?