zhenzhao / imas Goto Github PK

作者您好，感谢您出色的工作！在阅读paper时有处疑惑，在CVPR 2023 Open Access版本的paper中，3.3.1节中instance-based augmentations章节中的公式(7)：

是否应该修改为：

即，若当前样本较为困难（对应的$\gamma$较大）时，对应需要为弱增强版本分配更大的比例来弱化增强程度，期待您的解惑！

Hi! I see you conduct the ablation experiments of Intensity-based augmentations and CutMix-based augmentations. But I wonder if there still exists obvious improvement between vanilla CutMix strategy and designed CutMix-based augmentation strategy, including your another work AugSeg accepted by CVPR2023? Since when I conduct my own experiment, I find that just applying vanilla CutMix can yield a satisfying performance. Thanks!

作者你好，谢谢你提供这个好方法
发现你提及曾研究过UniMatch框架，iMAS本身是应用在教师-学生模型，但UniMatch没有引入教师模型
有点疑惑iMAS能否应用在UniMatch的框架上?
希望得到你的回覆，谢谢 !

[您好！我在实现自己的代码时，也遇到了U2PL一样的问题（详见此链接），即如果把教师模型设置为eval模式，掉点会很严重，所以不得不设置为train模式。但是我看到您这份代码把教师模型设置为eval模式，performance也很不错。想问一下您觉得把教师模型设置为eval模式会掉点的原因是什么？

Hi! Sorry to bother you here, but I just sent an email to you yesterday and I'm not sure if you have received it. If it's not too much trouble and you happen to see it, could you kindly drop me a reply? Thank you very much!

作者您好，请问我在超算平台上运行程序，申请了多张卡。但是只使用了一张卡，其它卡全部未使用，请问是什么原因，怎么能解决。

Hi! In Table 4 in your paper, there are two kinds of results. The bottom line of results corresponds to the output_stride of 8 (mIoU of 75.2 78.0 78.2 80.2 vs 1/16 1/8 1/4 1/2, respectively). And I wonder what's the number of output_stride in the other experiment(mIoU of 74.3 77.4 78.1 79.3 vs 1/16 1/8 1/4 1/2, respectively)?

在进行硬度定量分析的时候，问什么计算出的wIou还要乘上那个高置信度比例再算

When I tried learning using config_semi.yaml of citys_semi372, the accuracy was Best-STU:70.40/239 Best-EMA: 70.33/239, and the author's log Best-STU:76.94/236 Best-EMA: 77.43/188. There was a large discrepancy.
Due to the GPU, the number of training batches was set to 8 on a single GPU, so there is a difference in the learning environment, but is there really that much of a difference?

您好！请问您在代码里用dist.all_gather这个函数的时候，会出现以下几个问题吗？（1）显存利用率卡在100%，训练进程卡住；（2）训练不稳定，中途可能会报nccl相关的bug. 谢谢！

您好，请问您能提供voc的config文件吗，我自己改的不确定是否正确

你好，我看到代码中，关于无标签数据集，仅是对该类数据进行强弱增强，然后利用原有标签进行损失计算对吗？但是随机裁剪操作会不会影响数据与标签的对应关系，进而影响实验结果呢？

您好！请问您这份工作是在多少张什么类型的GPU上进行的？另外我看到您对cityscapes数据集进行推理的时候，采取的是sliding windows的方法。我自己用sliding windows推理（代码和UniMatch完全一致），在4张V100进行推理，需要大约3分钟，但是看您的日志只需要半分钟。想问一下有可能哪里出现了问题？谢谢！

您好，我想问一下在论文中的这个wIOU是您提出来的公式还是只是借用的已有的这个公式呢？

作者你好！在这行代码里，我看到你使用了dist.all_reduce，但是一般来说使用了DDP就不需要再把各GPU上的值给汇总起来了。请问这行代码的作用是什么呢？

Hello,

I'm trying to download the ResNet50 pretrained network listed in the ReadMe, but I am receiving this error when clicking the link:

"Sorry, the file you have requested does not exist."

Is there anywhere else I can find this file? I tried downloading resnet encoder weights from the pytorch repository but there seems to be differences between your code and that model file.

Thanks!

Hi! In Table 4 in your paper, there are two kinds of results. The bottom line of results corresponds to the output_stride of 8 (mIoU of 75.2 78.0 78.2 80.2 vs 1/16 1/8 1/4 1/2, respectively). And I wonder what's the number of output_stride in the other experiment(mIoU of 74.3 77.4 78.1 79.3 vs 1/16 1/8 1/4 1/2, respectively)?

您好，感谢你卓越的工作，我有一个问题是关于你在网络训练时使用分布式训练，那么原文中批处理是16，如果用8张卡的话，那么每张卡就2个无标签图片，那么增强就只能在两个图片上进行增强嘛？因为感觉每个gpu都只是处理自己的数据。

Hi! I just read the paper of iMAS and find the formula 8. Theoretical speaking, the harder the sample is, the larger its gamma is. So for the hard samples, it is inappropriate to assign large proportion of strong augmentation. For the easy samples, the strong augmentation can be added without too much concern. So I think the correct formula is written as: A_s <- (1 - gamma) * A_s + gamma * A_w. But in the paper it is A_s <- (1 - gamma) * A_w + gamma * A_s. So I am confused by it. Could you please explain it? Thanks a lot!

作者您好，我在运行您的代码的时候发现这个文件已经被删除了，请问您是把相关实现转移到了别的文件里吗，缺少这个文件对运行是否有影响呢

Hi! In iMAS the batch size is set as 16 + 16 = 32, while the batch size in most of the baseline works is set as 8 + 8 = 16. I wonder if it is a fair comparison since the segmentation task is sensitive to the bs?

I went to the pre-trained ResNet-50 and ResNet-101 links and it says the files do not exist.
Since we are interested in this paper and would like to verify it, would it be possible to upload the pre-trained model?

有试过不用easy-hard pair吗？就用labeled image，效果怎么样？我不明白为啥easy-hard的方法效果会好，能进一步解释一下吗？谢谢。