Open M-Chase opened 1 year ago
try teacher_loss_weights=1.0?
大佬跑通了吗,可以加一下联系方式吗?想有偿咨询一下我为啥没跑通
@Meoooww Under the custom project , we have optional Supervised training step, is it mandatory , how much data is needed in that step ? Last step is Directly starting of SSOD training, is it the 4th step in series of step ?what is the problem i idnt understand exactly. Thanks in advance
@Meoooww Under the custom project , we have optional Supervised training step, is it mandatory , how much data is needed in that step ? Last step is Directly starting of SSOD training, is it the 4th step in series of step ?what is the problem i idnt understand exactly. Thanks in advance
Sorry for the late reply. my question is how to adjust the parameters in yolov5l_custon_ssod.yaml: I used this model on my own dataset, and the AP of fully supervised Yolov5L is around 80, with slight overfitting. After semi-supervised on datasets of the same distribution, there is a serious overfitting problem. The rendered pseudo-label is not quite as expected (a large number of wrong labels and missing labels)
Here are a few parameters I considered, if there are errors or omissions, please point them out: nms_conf_thres, nms_iou_thres: to remove redundant prediction box ignore_thres_low, ignore_thres_high: Corresponds to τ1, τ2 in the paper, which is used to determine whether the pseudo label is reliable How do you regulate ignore_thres_low here, ignore_thres_high? My experiments have not yielded the desired results, it troubles me for a long time
Hi @Meoooww ,
I am having similar issues on some custom datasets - did you fix your problem by changing/tuning the parameters your mentioned:
Thank you :)
Hi @Meoooww ,
I am having similar issues on some custom datasets - did you fix your problem by changing/tuning the parameters your mentioned:
- nms_conf_thres, nms_iou_thres
- ignore_thres_low, ignore_thres_high
Thank you :)
Hi The paper did not mention how to adjust the four parameters, so when I failed to reproduce, I guessed that these four parameters needed to be adjusted. But, I didn't succeed, I don't know what principles to follow to adjust the four parameters. If you find a suitable tuning idea, please let me know as well
(Blindly adjusting parameters has taken me too much time, and I have temporarily abandoned this project)
如题,在自己数据集上训练的,这种情况是需要修改教师Loss的占比权重吗?