Surprisingly even with 1 example per class, results better than previously possible with 25 (before MixMatch) are achievable. On CIFAR10, with a single example per class FixMatch obtains between 48.58% and 85.32% test accuracy with a median of 64.28%.
-
-
Prikaži ovu nit
-
Code is up: https://github.com/google-research/fixmatch … And being my usual distracted self, I forgot one co-author from the list:
@alexey2004 (Sorry Alex!) The code for ImageNet will come later.Prikaži ovu nit
Kraj razgovora
Novi razgovor -
-
-
What happens when unlabeled data is unbalanced?
-
STL-10 is a dataset where the unlabeled data is noisy and labeled distribution is suspected to be mismatched (not knowing the labels, I can't tell definitely). SVHN class distribution is non uniform. Studying mismatched distributions is beyond the scope of this paper.
Kraj razgovora
Novi razgovor -
-
-
2a) what's the approx margin of pseudo-labeled examples kept? (approx margin e.g. https://arxiv.org/abs/1810.00113 )
-
We didn't look into that, I'm not familiar with this work. But the code is open source so hopefully you'll find it easy to experiment with / adapt to answer your question.
Kraj razgovora
Novi razgovor -
-
-
1) why do you think the method doesn't do better on cifar100? class similarity? can you disprove augmentation is to blame? (e.g. tiger gets augmented to jaguar) 2) can you quantify strong vs weak augmentation? perhaps using perceptual loss?
-
1. Given ReMixMatch does a bit better with fewer labels, I would suspect distributional alignment makes a difference. So my guess would be that distributional complexity is the limiting factor on Cifar100. I can't conjecture more with my current knowledge. Good question though.
- Još 3 druga odgovora
Novi razgovor -
-
-
Do you think something similar would also work for tabular data?
-
We haven't tried so I can't give a categorical answer. The method in itself only relies on weak and strong augmentation, a neural network and a loss. Finding the right augmentations for your tabular data seems to be only requirement to try the method.
Kraj razgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.