The performance of VAT algorithm is much lower.

Question

The performance of VAT algorithm is much lower.

LongLong-Jing opened this issue 5 years ago · 4 comments

Thanks for your implementation. I trained the VAT algorithm no the CIFAR10 datasets with the default parameters. However, it got 24.5% error rate on CIFAR10 dataset which is around 11% lower than the results that you reported in the table. Can I reproduce all the results in the table with default parameters?

LongLong-Jing commented 5 years ago

Thanks!

Answer 1 · 2019-07-10T01:30:43.000Z

Hi, LongLong-Jing!

All the results can be reproduced with default parameters.

You might have trained the model with CIFAR10 1k labels. (It means python build_dataset.py --dataset cifar10)
Did you run python build_dataset.py --dataset cifar10 --nlabels 4000
and python train.py --dataset cifar10 --alg VAT?

Answer 2 · 2019-07-10T01:49:18.000Z

Hi @perrying,

Thanks for your very quick response. Yes. I checked the parameters and found that I used 1k labels. How many labeled images should I use to reproduce the results in the table?

Answer 3 · 2019-07-10T01:59:37.000Z

You should use 4k labels for CIFAR10 and 1k labels for SVHN.