xiaoyao3302 / ccvc Goto Github PK

Hi! In your code the number of GPU is 4, and I wonder what kind of GPU did you use? Since the dual branches of model may bring lots of memory burden and I am a little worried about it. Thanks!

Hi！您好，您的工作非常出色和有效，但我在多次实验后有这么几个问题。我发现如果去掉loss_cosine，似乎在有标签数据越少的时候，子网越容易坍缩，从而导致性能很差，差不多跑三次就会有一次坍缩，导致效果差，但也不是每次都坍缩，可能这是与网络的随机初始化有关。在有标签越多的情况下，往往不是那么容易坍缩。我能理解这种现象是因为，有标签越多的情况下，学习到了更多的准确的语义信息，这样决策边界和预测会更加多样，从而子网越不容易坍缩。而loss_cosine则保证了下限，加上之后不管结果如何波动，都在可控范围之内，而不会出现很差很差的性能。请问我的理解是正确的吗，非常期待您的回复与指导交流

Excuse me, why does your unlabeled. txt file also have a mask image location corresponding to each unlabeled data？If I have a new batch of unlabeled data without corresponding mask images, how should I edit the unlabeled. txt file？

Hi! I wonder why i meet this error: argument --mode_confident: invalid choice: "'vote_threshold'" (choose from 'normal', 'soft', 'vote', 'vote_threshold', 'vote_soft', 'no').It seems that 'vote_threshold' appears to be a valid option.I'm eagerly awaiting your reply and greatly appreciate your amazing work. Thanks!

Hi, thanks for your good work, I have a question is why the CPS result for Cityscapes dataset is not mentioned in your Table3 ?

Hi! I have some doubt in your code~

1. In your code, the dilation in the experiment of Cityscapes is set as [12,24,36] and the replace_stride_with_dilation is [False, True, True], while other baselines set them as [6,12,18] and the replace_stride_with_dilation: [False, False, True]. So I wonder whether it will affect the results or not?
2. In your paper, the batch size of Cityscapes and Pascal are 8 and 24, respectively. But since you adopt the code of UniMatch, to the Cityscapes, I think the actual bs is 2 * 4(unlabeled images) + 2 * 4(labeled images) = 16, where 4 is the number of GPUs and 2 is the number of images per GPU. So I think the illustration of your paper is a little misleading~ How do you think?
3. Is the output_stride set as 16 in your ResNet backbone?
   Thanks a lot!

Hi! I wonder what the exact difference is among six settings of "mode_confident": ['normal', 'soft', 'vote', 'vote_threshold', 'vote_soft', 'no']. Looking forward to your reply and greatly appreciate your amazing work. Thanks!

Hi, thank you very much for your excellent work. I would like to ask you a few questions. When I was doing the ablation experiment, I found that if the unsupervised part of the loss is not added, there is only supervised loss, but the iou_ave is higher, but the problem here is that although the obtained iou_ave is higher , but the individual iou of the two branches is not high, and the value of both of them is very different from the average iou_ave, such as
[2023-08-29 18:01:01,619][ INFO] ***** Evaluation with branch 1 original ***** >>>> meanIOU: 58.85
[2023-08-29 18:01:01,619][ INFO ] ***** Evaluation with branch 2 original ***** >>>> meanIOU: 58.63
[2023-08-29 18:01:01,620][ INFO ] ***** Evaluation with two branches original ***** >>>> meanIOU: 60.27

But if I add the consistency loss,
[2023-08-29 14:20:26,233][ INFO ] ***** Evaluation with branch 1 original ***** >>>> meanIOU: 58.76
[2023-08-29 14:20:26,233][ INFO ] ***** Evaluation with branch 2 original ***** >>>> meanIOU: 58.73
[2023-08-29 14:20:26,233][ INFO ] ***** Evaluation with two branches original ***** >>>> meanIOU: 59.19

We can know that after the consistency loss is added below, the iou of the obtained branch is obviously higher, but the iou_ave is not only the unsupervised iou_ave. What is the reason for this? I guess it may be related to the way of evaluate. May I ask if you also encountered this problem during the experiment. Looking forward to your reply, much appreciated.

Sorry to bother you. Is there something wrong in Eq.4, 6, 7? If not, could you please give me some intuitions about these expressions?

Hi! The parameter "inplanes" is set as 128 in https://github.com/xiaoyao3302/CCVC/blob/master/model/backbone/resnet.py#L70, but in the original setting, it should be set as 64. I wonder why the difference exists. Thanks!

Hi! In this line of code, "center_crop" is adopted when training with evaluation. I know it is for speeding up the whole training process, but I wonder if there is difference between center_crop or sliding windows, and how large the difference is. Thanks!

Hi! After reading your paper, I think your CPL loss is basically the same as the one in CPS. However, you assign a bigger weight wc to the conflict area where one of the subnets are confident of its prediction. But I am not sure whether what I think is right or not. Thanks!