I have tried several times to reproduce the reported results in the Paper using the public code and config setting, but failed each time. I got results that were 7% worse than the reported number when I used train.txt for training and val.txt for validation.
Then I decided to run an interesting experiment that I used trainval.txt for training and val.txt for validation, I can reproduce the Paper number!
Followings are my trained models on train.val and trainlog.txt:
I have tried several times to reproduce the reported results in the Paper using the public code and config setting, but failed each time. I got results that were 7% worse than the reported number when I used train.txt for training and val.txt for validation.
Then I decided to run an interesting experiment that I used trainval.txt for training and val.txt for validation, I can reproduce the Paper number!
Followings are my trained models on train.val and trainlog.txt:
https://pan.baidu.com/s/1zhCt1P3KsVrCyBk8I8kkew https://pan.baidu.com/s/1P7IXetc1otXtDf0HJQ_7Ug
If any question, please discuss freely.