gank0078/acr

Pytorch implementation for "Towards Realistic Long-Tailed Semi-Supervised Learning: Consistency Is All You Need" (CVPR 2023)

acr's Issues

论文思路

作者你好！今天在CVPR官网看到你们的论文，深受启发！对于论文我有两个问题想请教作者：（1）引言部分所列举的三种分布不匹配情况（uniform/consistent/reverse）是怎么想到的？是通过其它论文受到启发还是在做实验的过程中发现这一现象呢？（2）论文对原始的logit adjusted loss的超参数τ进行了自适应改编，由于logit adjusted loss论文强调τ越大越关注少数类别，请问你们是怎么想到使用KL散度去自适应衡量τ的值呢？谢谢！

Why does standard branch use logit adjustment in the computation of L_con?

Greetings!
According to your paper, Eq (1) and (2) say that the standard branch follows the FixMatch training, and there is no logit adjustment at all.
However, according to your code, the standard branch actually use logit adjustment (args.adjustment_l1). I wonder why the inconsistence exists?

fails to reproduce

Hi,

Thanks for the nice work and open source code.

I cloned the code yesterday and ran the consistency setting in CIFAR-10 with the command provided in the README.md:

python train.py --dataset cifar10 --num-max 500 --num-max-u 4000 --arch wideresnet --batch-size 64 --lr 0.03 --seed 0 --imb-ratio-label 100 --imb-ratio-unlabel 100 --ema-u 0.99 --out out/cifar-10/N500_M4000/consistent

However, I failed to achieve similar performance as reported in the paper.

Here is my output:

current epoch: 500
05/19/2023 06:10:55 - INFO - main - top-1 acc: 74.44
05/19/2023 06:10:55 - INFO - main - top-5 acc: 97.74
05/19/2023 06:10:55 - INFO - main - Best top-1 acc: 76.66
05/19/2023 06:10:55 - INFO - main - Mean top-1 acc: 74.89

The reported performance in this setting of the paper is 81.6%. Is there something I missed?

I used a single GPU with PyTorch 1.4.

On how to reproduce the results on CIFAR100-LT

Hi,
Sorry to bother you again. I tried to reproduce the results reported in your paper on CIFAR100-LT. Here is my running: python train.py --dataset cifar100 --num-max 50 --num-max-u 400 --arch wideresnet --batch-size 64 --lr 0.03 --seed 0 --imb-ratio-label 20 --imb-ratio-unlabel 20 --ema-u 0.99 --out out/cifar-100/N50_M400/consistent
But I only reached an accuracy of around 44.5 (the reported on is 48.0). Is there something wrong of my settings?
Thank you~

Why use interleave and de_interleave? Why not just torch.cat?

Greetings,
I've been studying this wonderful work recently. I wonder why you use interleave and de_interleave functions in train.py? Can torch.cat do the same thing?

Dataset in train_split is not random

In tran_split function, the order of idxs seems fixed. Maybe you should set it random.

不太理解代码中提到的adjustment_l1和adjustment_l12的区别，感觉这两个值好像是一开始就设定好的一样

Question about the equation 3 and 4 in the paper

Hi! I have some questions about the equation 3 and 4:
(1) To equation 3: why the logit of balanced branch towards labeled data should minus the estimate of class prior? In the supplementary meterial, both of two branches use standard CE loss to labeled data.
(2) To equation 4: why the pseudo labels from the standard branch should add the estimate of class prior? I guess the latter is for generating more accurate pseudo labels, am I right?
Looking forward to your reply. Thanks!

Can you share the baseline implementation, too?

Hello. Thanks for the great work.

Besides, I was wondering if you can share your baseline implementation e.g., fixmatch.

Thank you.

Paper link is 404 error

Paper link is 404 error
what happened