ytaek-oh/daso

Repository for the paper `DASO: Distribution-Aware Semantics-Oriented Pseudo-label Imbalanced Semi-Supervised Learning'.

Hi there 👋

daso's Issues

FixMatch accuracy on CIFAR10-LT

Hi,

Thanks for the great work. I had a concern regarding the performance of FixMatch on CIFAR-10-LT in Table 1. With N1=1500, M1=3000, DARP reported 71.5% and 68.5% test accuracy for r=100 and 150, respectively. But your numbers are 77.5% and 72.4%, respectively. Could you please explain what is causing the performance difference here?

Thanks!

About the accuracy

Hi. Thanks for the great work here.

But I can't reproduce the result you reported in Table 2(Ours γu = 1/100 (reversed) N1 = 1500 M1 = 3000).I use the config in configs/cifar10/fixmatch_daso,and my command is python main.py --config-file configs/cifar10/fixmatch_daso.yaml \ DATASET.CIFAR10.NUM_LABELED_HEAD 1500 DATASET.CIFAR10.NUM_UNLABELED_HEAD 3000 DATASET.REVERSE_UL_DISTRIBUTION True

The result I get in result.json is around 77.9,three points lower than your result.

About the training time

Can I ask what is training time for an experiment?

Implementation of Semi-Ave dataset

It is a very wonderful work!

Can you release the implementation code of the Semi-Ave dataset?

Exact same wandb plots for STL-10 Crest/Crest+

For the same seed, even after changing the ALGORITHM.CREST.PROGRESSIVE_ALIGN argument to False (for Crest) the plots obtained for both Crest and Crest+ are the same. This happens only for the STL-10 dataset, and not for CIFAR 10/100.

Request to put code

Hello.

Can you please upload the code for the paper: - Distribution-Aware Semantics-Oriented Pseudo-label for Imbalanced Semi-Supervised Learning