How to reproduce performance in standard models?

[lucienne999]

Hi, the code I run according to the quick start in README.md.

python tools/train.py -n yolox-s -d 8 -b 64 --fp16 -o 

However, I get 39.6 mAP which is the performance of legacy models(also showed in paper) .

How to reproduce the 40.5 mAP? Should I update some args?

[hhaAndroid]

@LicharYuan The accuracy of my run is also 39.5.

[Joker316701882]

@LicharYuan @hhaAndroid Hi, the reported results are trained with batch size 128. We try batch size 64 and get the same results (39.5mAP) as you mentioned. However, the 40.5 mAP can be obtained under batch size 64 when you change the decay factor for the ema model from 0.9998 to 0.9999.

YOLOX/yolox/core/trainer.py

Line 166 in 15e8725

self.ema_model = ModelEMA(model, 0.9998)

We will soon make an official PR. If you want to achieve the reported with batch size 64, try to modify this line.

[hhaAndroid]

@Joker316701882 Thank you! I will try.

[RangiLyu]

@LicharYuan @hhaAndroid Hi, the reported results are trained with batch size 128. We try batch size 64 and get the same results (39.5mAP) as you mentioned. However, the 40.5 mAP can be obtained under batch size 64 when you change the decay factor for the ema model from 0.9998 to 0.9999.

YOLOX/yolox/core/trainer.py

Line 166 in 15e8725

self.ema_model = ModelEMA(model, 0.9998)

We will soon make an official PR. If you want to achieve the reported with batch size 64, try to modify this line.

Would you like to upload detailed training settings and logs together with the models? It would be greatly helpful for those who want to reproduce the results.

[Joker316701882]

@RangiLyu Make sense. We will upload the training log for yolox-s model. The only difference between the setting for this log and the current open-source code is the value of this decay factor (0.9999).

[Joker316701882]

For now, the training log for training yolox-s is available here:

Baidu Cloud:
url：https://pan.baidu.com/s/1DWzdVRPadLed0Z2nxdSg3g
password：gjk1

Google Drive:
https://drive.google.com/file/d/18AiPo8Bp-SacSEAyXVnZ9bVI9FtKa7Co/view?usp=sharing

[lucienne999]

For now, the training log for training yolox-s is available here:

Baidu Cloud: url：https://pan.baidu.com/s/1DWzdVRPadLed0Z2nxdSg3g password：gjk1

Google Drive: https://drive.google.com/file/d/18AiPo8Bp-SacSEAyXVnZ9bVI9FtKa7Co/view?usp=sharing

Thanks, I will try the decay factor = 0.9999 :).

[Joker316701882]

@LicharYuan Hi, 0.9998 is the current value. You need to change it to 0.9999 to reproduce our result.

[ajtvsv07]

Thanks for the logs @Joker316701882

can you please confirm that training has been done on 8 Nvidia V100 GPUs ?

[Xielihao]

@Joker316701882 Every version of YOLOX has to change 0.9998 to 0.9999？

[Joker316701882]

@Xielihao Yes.