perrying/realistic-ssl-evaluation-pytorch

Reimplementation of "Realistic Evaluation of Deep Semi-Supervised Learning Algorithms"

realistic-ssl-evaluation-pytorch's Issues

BN update

Hi @perrying
Thanks for sharing this code.
I saw that you have implemented the track_bn_statistic method in your network by overriding the foward method. This seems strange and what is the propose of using such way.

I noticed that track_running_stats=False does not work in this case.

The performance of VAT algorithm is much lower.

@perrying

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?

Some questions for your code

Hi, thanks for your codes. I have some small questions.
Q1. train.py: the 161th row: cls_loss = F.cross_entropy(outputs, target, reduction="none", ignore_index=-1).mean(). The outputs contains labled data and unabled data.
I think it may be inappropriate.
Q2. which way you use to calculate the loss in pseudo_label.py

some questions about mixmatch

Thank you for this repo,it help me a lot,but i run the mixmatch method,find that it does‘t have good performance. i want to know the reason abou this .i also tried some other hyperparameters ,but it isn't do well.