Nicholas Carlini
Nicholas Carlini
I'm not sure how to reproduce this. Can you clarify what you've done exactly step by step so I can try to help?
I've tried cloning with the latest set of dependencies and it's working for me. Do you still get this issue?
So right now I do think the collection has the possibility of going OOM when there are a lot of duplicates: it does load them all to memory at the...
The SVHN dataloader can be found here: https://github.com/google-research/fixmatch/blob/d4985a158065947dba803e626ee9a6721709c570/libml/data.py#L297-L300 We run experiments with the smaller svhn_noextra dataset here.
Yeah, we used the smaller dataset to train. This dataset, in the code, is called `svhn_noextra`. If you look at the ReMixMatch paper I think we also have results for...
That's interesting you find it's not working. We were able to reach ~state of the art accuracy at the time on semi supervised CIFAR-100 (with 100 classes) and ImageNet (with...
Sorry I don't really follow what you're saying here. You have 2000 classes, and with 400 images per class you get 48.8% accuracy. Is this with FixMatch? "If this is...
What accuracy does fully supervised training on these images give, ignoring the unlabeled data?
Supposing you don't use transfer learning, what accuracy do you get? The thing I'm trying to understand is this: supervised learning will strictly out-perform semi-supervised learning given the same number...
When we say in our paper that we apply a "weak augmentation" then this means (among other things): "flip, with 50% probability, the image across the vertical axis". When we...