-
论文中提到了一个损失Lda,这个损失加到有监督损失Ls上为什么就能缩小 有标签和无标签的数据分布呢?能帮忙详细描述一下吗?万分感谢!
-
我训练的时候同样遇到了KeyError: 'loss_cls'
,我因此在报错前面print了loss,发现在报错之前是正常的,但是报错的时候,loss少了loss_cls.
这是我print的地方:ssod/models/soft_teacher.py:243-245
loss = self.student.roi_head.bbox_head.loss(
…
-
![image](https://github.com/Adamdad/ConsistentTeacher/assets/56388226/4f7b8525-b747-467d-ae93-0a087e527549)
-
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?
-
Hi, when I reproduce your result, I find that I need 40min per epoch, is it normal? By the way, my device is a 3090Ti GPU with batchsize=32, 10% labeled coco
thanks for sharing
-
您好,之前看了您的论文受益匪浅,现在我想在**VOC**数据集上进行半监督训练,并尽量复现您在论文中的效果(如下所示):
,但与论文中的效果差了好多点,或许是我复现的过程和您的实现有出入,因此我有一些问题:
**1**、在使用**VOC**数据集进行训练时该使用哪一个配置文件,我使用的配置文件是源码中的**configs/ssod/voc/yolov5l_voc_burn.yaml**。但…
-
Does val set be used in training ? Or the val set just used to calculate the mAP? What is the difference of test set and val set here?
-
前辈好,我有两个问题想请教:
1.在设置debug=True后,存下来的图片一直都是空图(图片上并没有叠加伪标签),请问这个正常吗(COCO训练过程打印的mAP还挺高)?还是哪里没设置对?在设置debug为true后,我根据报错提示(提示存储路径不存在)修改了图片保存路径
2.我先试了COCO数据集,能正常运行。然后在自己的7类数据上进行SSOD训练报错,后来发现好像在加载预训练模型时,需要在…
-
iu110 updated
8 months ago
-
When I was developing a semi-supervised detection library, I followed your work closely. I was sorry to see that your paper was withdrawn from ICLR, but congratulations on getting it accepted at CVPR!…