search

LiheYoung / UniMatch

[CVPR 2023] Revisiting Weak-to-Strong Consistency in Semi-Supervised Semantic Segmentation
https://arxiv.org/abs/2208.09910
MIT License
478 stars 60 forks source link

Question about the training of fixmatch #54

Closed yasuobinggan closed 1 year ago

yasuobinggan commented 1 year ago

Hello, I experimented with unimatch and fixmatch separately under the condition "VOC2012 92labeled-GPU 4*3090". The code is not adjusted. I found that the training of unimatch is stable and can achieve very good performance, but the training of fixmatch is very unstable and fails to show a state of gradual convergence after climbing.

The validation set curve: The blue line is unimatch The yellow line is fixmatch

I would like to ask the instability of fixmatch is normal? Do you have any relevant solutions? Thank you so much! image

LiheYoung commented 1 year ago

According to the curve, it seems you tried the 92 labeled setting. Due to the extremely scarce label, I think such instability is acceptable for FixMatch. If I remember correctly, other settings with more labeled images will be much better and more stable, such as 732 and 1464.

yasuobinggan commented 1 year ago

Get it, thank you!