dnap512 / selfreg Goto Github PK

Hello,

I can't find the file or folder of “domainbed” in this repo.
Could you share it with me?

Thanks!

Thanks for the wonderful work

One thing I want to ask is the backbone network of your work.
The paper says that you used ResNet18 for the backbone network.
However, in DomainBed, all experiments are conducted with ResNet50.
You used the same results from DomainBed paper for other baselines (which used ResNet50) in your paper but your method is conducted with ResNet18.
Can you clarify the experiment setting regarding such issue?

Thanks, in advance

Hi, I have a question, in

Hi, thanks for your great work and code release.
I have a question about training.

It is noticed that in train.ipynb
lr_decay-epoch is set to 100, but max epoch is 30.
In paper, 'note that such a decaying learning rate is not used when ti combined with the Stochastic weights averaging technique'

It means that: if we use SWA(stochastic weights averaging), we use constant lr (0.004) during training.
if we not use SWA, we need to decay lr to 0.004*0.1=0.0004 after epoch 24.
Am I right ?
Another question is :
the results of these 2 settings differ a lot ?
Thanks for your help.

Hi, could you please provide the code of using SWA? Thanks.

Hello, when I read your code, there is problem.
In your code, SelfReg-main/codes/utils.py line 224-226
if is_self_reg: output, feat = model.extract_features(x) proj = model.projection(feat) elif is_positive_pair: output, feat = model.extract_features(x1) output2, feat2 = model.extract_features(x2)
I want to know the function about elif, what the is_positive_pair is and x1, x2 is what,
Wish your respond.