训练自己数据集时l_imgcls损失指标异常
Closed this issue · 5 comments
gen0924 commented
yuantn commented
你好,你可以调试一下代码,看看在 mmdet/models/dense_head/MIAOD_head.py
中 l_imgcls
这个函数的训练过程中每个变量的情况有没有异常。
You can debug it and check whether there are any abnormalities in the training of each variable in function l_imgcls
in mmdet/models/dense_head/MIAOD_head.py
.
gen0924 commented
yuantn commented
有可能是这样的吧,你可以试试在数据集中新建另一个类别,但那个类别没有对应的图像,测试一下是否还会出现这个问题。
Maybe it is right. You can try to create another class without any corresponding image in the dataset, and test whether the problem still occurs.
gen0924 commented
按照你的意思,我增加一个类别,但没有数据,i_imgcls不会为0,但随着训练的进行越来越小,这感觉是正常的趋势!所以,因为我前景只有一个类别,随意导致回归框类别损失一直为0?
yuantn commented
我感觉原因是这样的。
I think that the reason is like this.