如果训练断掉了 怎样从断点继续训练而不是从头开始训练

[Haochen-Wang409]

在 saver 下面添加 auto_resume: True 即可

[yjq767579182]

ok thanks！

[yjq767579182]

I still have a question, for the encoder, can I use any feature extractor, like some mainstream transformers, and I still use deeplab or other decoders

[yjq767579182]

for the apply_aug, wherther have other choose

[Haochen-Wang409]

Of course, you can use any other backbones. As for apply_aug, we support CutOut, CutMix, and ClassMix.

[yjq767579182]

我想请问这段代码是实现了论文里的这个功能吗

[Haochen-Wang409]

是的，对于每一个类别，我们都维护了一个 memory bank 去储存不可靠的伪标签

[yjq767579182]

oh! thanks for your answer！I'm very appreciated

[yjq767579182]

hello i'm here to ask a question again，I would like to ask if the 1 in the figure is set to 3, 4 or larger, will it have a better effect on semi-supervised，

[Haochen-Wang409]

You can try it. We did not ablate this.

[yjq767579182]

ok thanks

[yjq767579182]

hello，I would like to take the liberty to ask, ema_decay=0.99, so why do we need to set the teacher model and the student model? The specific gravity coefficient of 0.99 is not much different.

[yjq767579182]

Hello author, I still have a question to ask, why does the marked data still have a threshold, as shown in the picture

[Haochen-Wang409]

hello，I would like to take the liberty to ask, ema_decay=0.99, so why do we need to set the teacher model and the student model? The specific gravity coefficient of 0.99 is not much different.

The teacher model is used to generate pseudo-labels to supervise the student model for each unlabeled sample.

[Haochen-Wang409]

Hello author, I still have a question to ask, why does the marked data still have a threshold, as shown in the picture

Features of those low confident labeled pixels might be biased.