kinux98 / ssl_eln Goto Github PK

您好，我想问一下如果要跑自己的数据集，unlabeled图片是不是也需要像增强集那样生成轮廓的粗标注？我在run的时候发现unsample会同时输入原图以及SegmentationClassAug的图。期待您的回复。

Hi! In your paper, the embedding fi in the contrastive loss comes from the student model. But in this line of code, the dec2 is a part of the teacher model, thus the feature_vec_un_k comes from the teacher model. Is the code contrary to the statement in the paper? Looking forward to your reply, thanks a lot!

How to train the ELN network, the first training stage for ELN is not found.

hi，Congratulations on your outstanding work. I wonder if this is reasonable that when computing contrastiveloss the query features come from teacher model while key features come from students model?
code : pxl_dist_aug = pixelwisecontrastiveloss(self, feature_vec_un_k.detach(), feature_vec_un_aug, final_candid, final_indices)

Hi! I just found the file "requirements.txt" is filled with lots of package ending with "@ file: ", and I wonder whether you export the file in a wrong format or not. Maybe the command “pip list --format=freeze > requirements.txt” is more suitable than "pip freeze > requirements.txt" ? Thanks!

How to train the ELN network, the first training stage for ELN is not found.

Hello, thank you very much for your excellent contribution, I have a question, how many GPUs did you use when you did the experiment, and what type of GPU was it.

Thanks a lot for your good work! And I have a question about performance comparion. In another work called CPS, The MT model they compared has better performance than this article in same network architecture(deeplabv3+ with resnet50). What causes this?

hi,
in the ‘train_both’ process，when the epoch_number is bigger than the warm_up_number, enc2 and dec2 will import enc1 and dec1 parameters. Code is as following
if epoch >= (self.warm_up_epoch+1):
self.enc.load_state_dict(self.enc2.state_dict())
self.dec.load_state_dict(self.dec2.state_dict())

Will the exponential moving average (EMA) method make the parameters of the student model and the teacher model consistent at the beginning of each epoch, or is this paper deliberately doing this?

Hi, first of all thanks for your sharing. In addition, there is another question: Can you provide a tutorial for the installation of the environment? The installation environment provided in the project always has problems, and the provided requirements.txt file has many non-existing versions.

Hi! I find that in this line of code, there is a variable named iter_per_epoch = 266. May I ask how the result 266 is calculated? Thanks!