xyzach / rlsss Goto Github PK

Hi authors, thank you for the amazing work!
I am using your repo to train on another dataset (OCT images) but the results are quite bad. The student model cannot learn anything from the teacher model and performs very badly. Besides it, the teacher model seem to be affected by the student model and eventually performs worse than the initial teacher model. I used U-Net for this experiment.
Do you experience this and have any reasons?
Thank you.

Figure 1. Initial results of teacher and student models

Figure 2. Results after 100 epochs.

Hi,

Could you please explain what's the difference between your approach and meta-pseudo label (https://openaccess.thecvf.com/content/CVPR2021/papers/Pham_Meta_Pseudo_Labels_CVPR_2021_paper.pdf), as I don't find the related reference appear in the paper?

Cheers

Dear authors,
Firstly, thank you very much for the amazing work.
I try to reproduce your work but it does not work. In your paper, it's reported that you achieved 90.06 Dice, 82.01 Jaccard, 2.13 ASD, and 6.70 95HD. But when I use your code and retrain, I only achieve 87.82 Dice, 78.50 Jaccard, 2.35 ASD, and 10.23 95HD. I use the pretrained VNet model from UA-MT repo .
Can you provide some information about the environment that you trained the RLSSS on or any other settings in your code?
Thank you very much