Dawn-JN
commented
2 years ago I notice that in paper "...with batch size of 48....The initial learning rate is 0.001",but in yolox_s_mix_det.py, line 29
self.basic_lr_per_img = 0.001 / 64.0
The learning rate here is divided by 64, it's the same as YOLOX, but YOLOX's batchesize is 64 and bytetrack's is 48.
And in before_train function(trainer.py, line 154),
self.lr_scheduler = self.exp.get_lr_scheduler( self.exp.basic_lr_per_img * self.args.batch_size, self.max_iter )
So should it be replaced by 48 here?
Hello! I tried to train bytetrack_s_mot17(YOLOXs) twice, and both time the AP is 0.593(MOT17 Train), however AP=0.639 when using the provided weight.I train and test like this (with 8 K80 GPUs):
python tools/train.py -f exps/example/mot/yolox_s_mix_det.py -d 8 -b 48 --fp16 -o -c pretrained/yolox_s.pth python tools/track.py -f exps/example/mot/yolox_s_mix_det.py -c pretrained/trained.pth.tar -b 1 -d 1 --fp16 --fuseTrain on CrowdHuman, MOT17, Cityperson and ETHZ, evaluate on MOT17 train.Is there anything different or needs to be changed when training bytetrack_s_mot17(YOLOXs)? Please help me, thanks a lot!@ifzhang