yyzharry / multi-domain-imbalance Goto Github PK

In the file 'mdlt/scripts/download.py', there is a file "scripts/misc/domain_net_duplicates.txt" required. Can you provide this file?

Which among the three model selection methods proposed in the DomainBed paper did you use?

Does this repository has the code used for experimenting and creating Table 9,i.e., experiments related to Domain Generalization?

Hi! Thanks for the great work! I was wondering how is the zero-shot domain-class pair's mean representation $\mu_{d, c}$ calculated. Was not able to find details in the paper. Thanks!

Hello,
I am pleased to read your paper published at ECCV 2022. We noticed that in multi-domain environment, the number of training samples contained in each domain/environment varies significantly, and the parameter 'batch_size' means the number of samples sampled from each domain. So, how is parameter 'CHECKPOINT_FREQ' (i.e the parameter Epoch e in the pseudo-code ) set? It depends on the largest dataset or the smallest dataset? Can you tell us something about your experience?

Thank you very much!

Hi,

I am trying to replicate the BoDa loss with Mahalanobis Distance, however, the covariances of the features for each class/domain that I calculate have been singular and therefore non-invertible.

Is this a problem that you have encountered before and are there any easy solutions?

Thanks

Hi. Is there any update as to when you will be releasing code? Thanks.