thuml / self-tuning Goto Github PK

No module or files named 'efficientnet_pytorch' in the whole file holder. But 'from efficientnet_pytorch import EfficientNet' appears in the first line of efficientnet.py
Is there any thing should be done before running main.py? Thanks for your reply!

code snippet of models.py( line 121-122):
PGC_labels = torch.zeros([batch_size, 1 + self.queue_size*self.class_num]).cuda() PGC_labels[:,0:self.queue_size+1].fill_(1.0/(self.queue_size+1))

Hi, I can't connect the "https://cloud.tsinghua.edu.cn/f/04356d49d0054092b07e/?dl=1" to download data of Aircraft. I wonder if there are some connection problems of tsinghua cloud.

Links for all checkpoints and data are invalid

You mention that you use Efficientnet for cifar100 in your paper.

Your corresponding code is like

Can Cifar100 uses the imagenet-pretrained model(efficientnet)? The resolution differs a lot.

I really appreciate it if you could answer my question!

其他半监督学习方法如fixmatch只会使用阈值大于0.95的预测值生成伪标签，但是self-tuing直接照单全收了，这是不是不太合理

Hi,

Thanks for the awesome work and the public repo.

I wonder if it is possible to release the codes and training hyperparams of compared baseline methods in the paper (e.g., Fine-tuning, pseudo-labeling, fixmatch, etc.). I believe the further open source codebase will help the community to do more explorations and bring your paper more citations and impacts.

Cheers

Hi,

Thanks for the interesting work and sharing the code.

Recently, I reproduced the Fine-Tuning baseline method based on the released code for Self-Tuning method (directly delete the unlabeled and contrastive parts and use the same optim hyperparam and schedule), and the reproduced results are as follows (all experiments are conducted on 15% label proportion setting):

As the table shown, there is a huge performance gap between the reported numbers and the reporduced ones. Furthermore, I also found some reproduced numbers even much better than the reported numbers of SSL methods. As shown in the following table, the performance gap is quite unreasonable since large amount of unlabeled samples have been further utilized in these SSL methods.

So, I am really wondering how do you train the baseline methods to get the reported numbers?

Hi,

After reading ur paper and code, I found the PGC loss u implemented is a little bit different from the Eq (4) in ur paper? (u use KLD in ur code but not mentioned in the paper) Am I right, or I missed something?

The video in ICML 2021 website is not available for the non-registered users.

So it will be very helpful if you can upload the video on bilibili, thanks!