microsoft / Semi-supervised-learning

A Unified Semi-Supervised Learning Codebase (NeurIPS'22)
https://usb.readthedocs.io
MIT License
1.32k stars 176 forks source link

Use softmatch for semi-supervised image segmentation #90

Closed Fyw1988 closed 1 year ago

Fyw1988 commented 1 year ago

Hello, I tried to use your softmatch for semi-supervised image segmentation, treating each pixel as a dichotomy, but found that the mask calculated for each pixel is a fixed value except 1 (0.9999998807907104). There will not be significantly more masks less than 1 as the training cifar dataset does with iteration. Is there any adjustment if you want to use softmatch for semi-supervised segmentation?

Hhhhhhao commented 1 year ago

I'm not quite familiar with semi-supervised image segmentation. Can you specify more details? Or you can add my wechat account so we can discuss more on that.

Fyw1988 commented 1 year ago

I'm not quite familiar with semi-supervised image segmentation. Can you specify more details? Or you can add my wechat account so we can discuss more on that.

Thank you very much for your work! I've seen what went wrong

yuyang16101066 commented 1 year ago

I'm not quite familiar with semi-supervised image segmentation. Can you specify more details? Or you can add my wechat account so we can discuss more on that.

Hello, I have a question about reproduction of softmatch in cifar100 2500 labeled setting. I use wrn_28_8 as backbone, but can only get accuracy of 0.67. Do you have any idea about my question?

Hhhhhhao commented 1 year ago

I'm not quite familiar with semi-supervised image segmentation. Can you specify more details? Or you can add my wechat account so we can discuss more on that.

Hello, I have a question about reproduction of softmatch in cifar100 2500 labeled setting. I use wrn_28_8 as backbone, but can only get accuracy of 0.67. Do you have any idea about my question?

The result you reported here is very similar to what we got from wrn_28_2. So maybe check the config file and training log about the backbone? You can send the log file here.