基准(faster_rcnn_r50_fpn_1x_coco简称frrf)
config mAP(最好) 模型大小 publish_model cal_train_time (s/iter)
1.frrf 0.9346 315.32MB 158.23MB 0.1298
2.frrf+fp16 0.9301 236.47MB 79.38MB 0.0947
3.frrf+fp16+softnms 0.93526 236.47MB 79.38MB 0.0973
4.frrf+fp16+GIOULoss 0.9367 236.47MB 79.38MB 0.1076
5.frrf+fp16+GIOULoss+softnms 0.9319 236.47MB 79.38MB 0.1031