bit-da / rotta Goto Github PK

作者你好，我下载好预训练模型之后，重新运行代码还是会自动下载预训练模型，然后导致代码运行失败，请问这是什么原因造成的呢，以下是我文件存放的形式

Hi Longhui,
Thanks for your released code of this excellent work! I like this work as it sounds great.

During I was reading the code, and I found one minor error here: https://github.com/BIT-DA/RoTTA/blob/main/core/data/datasets/common_corruption.py#L22

Under my understanding, here may be:

d_id = i_s * len(self.corruptions) + i_c

, otherwise, the d_id will have duplicate values. Please check again here.

Yongyi.

Hey Longhui,

I am happy to see you release the code of CIFAR10 and CIFAR100 datasets, but there is still a little pity that I cannot see the implementation in DomainNet dataset, which is mentioned in your paper. Could you release the code of DomainNet dataset soon? Or could you tell me how to implement it based on the existing code.

Best.
Yongyi

Hi authors,
Thanks for your code! My question is:
Is the size of memory bank always be 64 and keep the same with bsz?
If so, when the number of categories are large (like CIFAR-100-C or ImageNet-C), CSTU would frequently discard previous data (according to Algorithm1 line5-9), and is it randomly discard previous data?

From what I understand, the ROTTA method seems to have a teacher-student structure like the COTTA method, and weight updates should occur not only for batchnorm weights but also for the overall weights. However, in the actual implementation, such as in the configure_model function, only batchnorm weights are updated. Am I correct in understanding it according to the implementation?

作者你好，
ImportError: cannot import name 'Conv2dNormActivation' from 'torchvision.ops.misc'我在运行代码时会出现这个错误，但是我的torchvision版本和项目中给出的完全一致，非常感谢作者的解答

Hello, I was wondering what is the purpose of target_mean and target_var line 18 and 19 in the bn_layers.py file: https://github.com/BIT-DA/RoTTA/blob/67e34c900cdd355fc07e55edd4c577ea7b8ebcc9/core/utils/bn_layers.py#L18

It seems that these variables are not used. Same for current_mu and current_sigma.

Additionally could you give more details on how the behavior of the Robust Batch Norm differs from the one of the normal Batch Norm (which also update its stats with moving average).

Great work by the way!