设置with_da_loss为True后，精度不如默认False高，请问这是为什么呢？

[BowieHsu]

@beauty-girl-cxy 您好， 有具体的log能提供吗

[LittleRain626]

@beauty-girl-cxy 您好， 有具体的log能提供吗 results.csv 您好，感谢您的回复，这个文档是仅修改with_da_loss为True后在coco的一个子数据集labeled-1%上的结果，很抱歉修改之前的我找不到了，根据之前我统计的结果，with_da_loss为False时map50为0.017。

results_.csv results_.csv是with_da_loss为False在coco子数据集labeled-2%上的结果。

[BowieHsu]

@beauty-girl-cxy 您好，我看了log，好像mAP50是0.042，所以打开了da_loss之后，精度会下降多少呢

[LittleRain626]

@beauty-girl-cxy 您好，我看了log，好像mAP50是0.042，所以打开了da_loss之后，精度会下降多少呢

result.csv是打卡da_loss后的结果，精度几乎变为了0

[jaideep11061982]

@BowieHsu how you interpret the results at the end to know how well framework has labelled unlabelled data ?

[Meoooww]

@BowieHsu how you interpret the results at the end to know how well framework has labelled unlabelled data ?

i see your question under many issue in my opinion, since the data is unlabeled, you cant evaluate model on it. bt u can use the unlabeled data to train, and evaluate model on other labeled data