-
Notifications
You must be signed in to change notification settings - Fork 174
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
How's FixMatch compared to supervised methods? #52
Comments
I'm not quite sure I follow the question. Do you want to know how well our methods do in comparison to a fully supervised baseline that uses, say, just 40 or just 250 images? This is given in Table 9 and Table 10. If instead you mean fully supervised on the entire dataset, then these numbers are reported in our ReMixMatch paper Appendix Section 4.1 To begin, we train a fully-supervised baseline to measure the highest |
Actually I meant both. Thanks for the considerate response. I implemented FixMatch with RandAugment in my own codebase and on a classification task, with real world data and imbalanced distribution (4 classes, majority:minority up to 50:1). The training dataset is composed of balanced/imbalanced labeled data and imbalanced unlabeled data (sampled randomly from the entire dataset). I tried a coarse grid search for hyperparameters (weight decay, tao, lr, lambda_u, range of augmentations), but got no luck. FixMatch didn't really help to improve classification performance compared to a purely supervised model. I have a few observations:
Here are my guesses, please correct me if I'm wrong. First, FixMatch should start the training process with a good or at least mediocre pre-trained model to prevent low quality pseudo labels introducing too much noise to the learning process. Second, FixMatch seems to have problems adapting to imbalanced classification tasks. The majority class is quickly overfitted. The randomly sampled unlabeled data cannot provide much useful info via consistency regularization and pseudo labeling. Third, hyperparameter tuning (maybe among all, augmentation range tuning is the most important) is vital and difficult in adaptation. I appreciate any comment or suggestion for improvements. Thank you very much. |
Yeah, unbalanced datasets are so far something that SSL really struggles with. We've found distribution matching from ReMixMatch can help here, but I don't know of any answers. My guess is this is less of any issue with fixmatch-as-expected, and more of an open research question for how to solve unaligned data distributions better. |
Thanks for sharing the great work.
It might be interesting to compare FixMatch with supervised learning methods, given the same labeled data and model structure, to further demonstrate the power of SSL.
I appreciate if you could comment on the idea or release experimental results on such comparison.
Regards,
The text was updated successfully, but these errors were encountered: