Loss rise problem

[Jackson-zhao97]

I use the voc data set for training. First of all, I hope that the teacher can get the pseudo label as accurate as possible at the beginning, so I have done enough training to make the model converge on the labeled data. However, when I started training in the semi-supervised part, I found that the loss continued to rise, and the predicted mAP did not change. Is it because the pseudo label introduces more noise?

Another problem is that in the pseudo label, there are often cases where the prediction is empty. In this case, my processing is to let the classification label be 0 (background class), and the bbox coordinates are [0, 0, 0, 0]. Is this feasible? Or is there another way?

I reproduced it in mm-detection, so some questions appeared. I hope to get your answers.

[Jackson-zhao97]

I found that setting bbox to [0,0,0,0] will cause the loss to be nan. But how can I deal with this kind of problem? If there is no bbox with a score greater than the threshold, how should the loss be calculated?

[OliviaZhang1996]

Maybe you can remove the unlabelled images with empty predictions. And just use the unlabelled data with predictions.