salesforce / comatch Goto Github PK

Hi! Have you conducted the qualitative ablation study of the designed components, including smoothed pseudo labels (equation 8), graph-based contrastive learning (equation11) and the memory bank? It seems that you only conduct the hyper-parameters ablation study. Thanks!

Thanks for sharing your clean codes.
While I find that the codes of distribution alignment (DA) only calculate:
q = Norm(q/mean(p(y))),
without multiplying labeled data distribution p(y) and further sharpening, which seems not same as the DA introduced in ReMixMatch. Wonder why don't use the standard DA to train models? Is the simple way you used better for SSL models' performance?

Hi,
Thank you for your great work, I am new to this area, I have the following questions:
1、How much improvements can be made by the branch of Graph-based contrastive learning? I can not see that ablation experiments in the paper. I'm interested in this part and would also like to know the performance improvements that might be useful for my task.
2、Does the Co-match method fit for multi-label classification tasks？

Once again, thank you for your nice work and clean code! Looking forward to your reply.
Best regards!
Tan

the idea in your paper is amazing，great truths are all simple. I have the following questions:
1、Does a stronger data enhancement method, such as RandomAugment, improve the performance?
2、To improve the performance, is necessary expend the size of distribution alignment ?
3、how to adjust the memory bank size when more label data is used in ImageNet ,such as 20% of ImageNet?
4、If use 20% of ImageNet data as a label, what are the recommendations for other hyperparameters?
5、From your point of view, what are the main challenges in achieving full oversight with 20% of ImageNet data?

Hi, thanks for the nice code.
Could you please provide some information on where can I obtain the public mocoV2 pre-trained model.

Dear Authors

Your work is exciting! I am trying out your code with the example you provides, python Train_CoMatch.py --n-labeled 40 --seed 1
I am running on one A100 gpu. I found that the speed is very slow, takes ~10 minutes for 1 epoch. Is this the expected behavior? I just want to make sure i am running the code correctly.

Thank you, and hope to hear back from you!

Dear Author
Thank you for this exciting work!
I have a clarification question regarding the experiment setting: Did you use any validation data to tune the hyperparameters for CoMatch? How did u choose the alpha, lambda etc?

Dear authors,
In Eq.10, you build the embedding graph with two strong augmentations, where the off-diagonal values are generated by enhancement 1 only. However, in your code, this is not implemented, this seems at odds with the paper description, could you explain it?

Thanks a lot for sharing your clear codes. I try to train the models on STL dataset but don't find the dataloader of STL suitable for SSL setting. Would you like to share the codes of STL Dataloader with randomly splitting labeled&unlabeled data?

Hello, I can't find the implementation of svhn dataset in this project, do you have open-source the code of this part? thank you!