Gang Li

@liuliu66 I have encountered the same problem. Have you solved it? thank you!

> Maybe there is something wrong with your Pytorch version. Could you please share us with your Pytorch version? I use Python2.7 and Pytorch1.0.1, and can't achieve reported results.

> @LIszu the paper is available: https://arxiv.org/abs/2002.09053 Thank you for your sharing.

Proposals是由teacher model的RPN产生的。在student的训练过程中，虽然student RPN会产生自己的proposals，但是没有被使用。而是把teacher的proposals直接拿来训练student了。

Which version of mmdetection do you use? The attribute "use sigmoid" is suppressed in latest mmdetection, which may account for this error. A suitable mmdetection is already placed in [PseCo/thirdparty/mmdetection](https://github.com/ligang-cs/PseCo/tree/master/thirdparty/mmdetection)....

Sorry for late reply. You need to modify the code in Line 206 of [train.py](https://github.com/ligang-cs/PseCo/blob/33b06eeebaf456213f7c2b0cb25122b81fa50897/ssod/apis/train.py#L206). Please use: `runner.load_checkpoint(cfg.load_from, revise_keys=[(r'^', 'student.')])`

Label-level consistency learning can be regarded as vanilla multi-scale training.

view 2是对view 1进行2倍下采样得到的，是低分辨率的图片。而检测器对低分辨率的图片，检测性能向来较差。通过Feature-level的对齐，因为view 2和P2-P6 feature分辨率和view 1的P3-P7 feature分辨率是一样的，可以使用相同的pseudo box来同时监督view 2和view 1的FPN features，我们希望低分辨率图片(view 2)，可以产生和高分辨率图片(view 1)一样强的FPN features. 具体的实现麻烦参考代码：[PseCo_frcnn.py](https://github.com/ligang-cs/PseCo/blob/master/ssod/models/PseCo_frcnn.py).

> Hello author, the above error occurs when I execute the training script, what is the reason? Which mmdet do you use? mmdet in the folder [thirdparty/mmdetection](https://github.com/ligang-cs/PseCo/tree/master/thirdparty/mmdetection) is prefered.