najaemin92 / fixbi Goto Github PK

Hi,

Thank you for releasing the code.

I run the code on the Office-Home dataset with the default parameters, but I find the accuracy is very low. I also try to tune the parameters, but it seems does not work. Can you provide the parameters that you use for the Office-Home dataset? Thank you.

Hi @NaJaeMin92,
I tried using the repo but I was unable to reproduce the results office dataset. My SDM accuracy got decreased to 51% and TDM is stagnant at 6%.

Hello, can you give me the pretrain weight of DANN on Office, OfficeHome, and VisDA, respectively?

Hi, do you have any suggestions on how can we choose warmup epochs for custom datasets?

Also, from the following line, you seem to be switching off the self-penalization loss after few epochs when other losses start to act. Why is it so?

The CVPR 21 conference is already finished this week. Could you share your code so that others can re-produce your results & build upon your work?

Hello,
Thanks for your great work. I have a quesion regarded to the following paragraph.

In your paper, the negative pseudo-label indicates the most confident label (top-1 label) predicted by the network with a confidence lower than the threshold τ. However, it may seem counterintuitive that the negative pseudo-label is unlikely to be correct. Despite having a low confidence value, it still possesses the highest prediction value among all categories.
I'm feeling a bit puzzled; could you provide me with some guidance or instructions?

i'm trying to transfer dataset that has labels to simular dataset without labels
like in this example
https://www.youtube.com/watch?v=9d4jmPmTWmc&t=570s

from the code that load the datasets its look like both the source and target dataset supposed to have labels.

Hi, according to your paper, the threshold \tau should be calculated as (mean - 2 * std)

Thus, in here:

It seems to be:
threshold = top_mean - args.th * top_std

Is my understanding correct? Please help me check about it, thanks!

Hello, I find that all of the total epoches, bim_start, sp_start and cr_start are equal to 100. So Bidirectional Matching and Consistency Regularization will not start in this case. Can you check the code and reply me?

Hi, thank you for sharing work!

I am a bit confusing about why top_prob_sd is used instead of top_prob_td in below code.

Thank you!

As for self-penalization with negative pseudo-labels, in the paper, it's not limited by the epoch. However, in the code, it just runs when the epoch < sp_start.

Can you provide your code?

@NaJaeMin92
Hi, I read FixBi interestingly and want to reproduce some results.

Can you provide .pt files?

You mentioned this repo for pretrained weights, but to me, it seems two repo has different network architecture (so cannot be loaded even if I train using the mentioned repo).

Should I edit networks' architecture of DANN (as same as this repo), train it on the Office-31 dataset, and use it as a pretrained baseline weight?

Thank you for your work and I will wait for your response.

need an answer: why these pretrained baseline model's weights cannot be provide?