I follow the default setting of codes and obtained 71.687% mIoU when training with 1/8 labels on pascal voc，but this paper reported 73.67% mIoU. Why the performance gap so huge, and anyone can tell me how to improve this code? Thanks very much!

Hi, Xiaokang, Thanks for sharing so solid work! I noticed that the loss in supervised learning is ohem loss. Have you ever done the experiments for ce loss and how about the result?

When I use pytorch 1.0, apex DDP will automatically exit without error message, so I try to run this code with pytorch 1.8.1. However I got similar results reported in #14 (72.0±0.2 mIU for cps, 1/8 voc r50 setting), can anyone figure out the possible reason that causes the performance decay? @yh-pengtu, @frank-xwang

Thanks for sharing this excellent work!

I have a question for the parameter initialization in this Teacher-Student paradigm:

• At the very beginning of training, the parameters for the Teacher net are initialized by that of the Student net, as in https://github.com/charlesCXK/TorchSemiSeg/blob/main/exp.voc/voc8.res50v3%2B.MeanTeacher/network.py#L30.
• Under this case, will the copy destroy the original setting? The parameters (of the semantic segmentation head) for the Teacher and Student are initialized differently as you said, but they are now unified by param_t.data.copy_(param_s.data).

Looking forward to your clarification! Thank you

Hi.According to the Getting Started.md, there is different epochs for different label ratio. but i'm confused about the setting's meaning. if i'm right, here you only need to specify the total iters.It is not necessary to calculate the epoch corresponding to different label ratio?

If I want to train supervised baseline, does it mean that if my total iters are 40k and the label ratio is 0.125, then I need to do 40k iterations(maybe batch size is 8) on this 0.125*10852 labeled images only? But if i use semi supervised method, one iteration means fetching the labeled data once and fetching the unlabeled data once?

Hope to receive your reply. very thanks.

best,

Hi again, I have more question about BN layer.

I want to disable batchnormalization layer, so I tried to change model status from model.train() to model.eval() in train.py. (And also tried replacing BatchNorm layer with Identity layer but same error occurred)

The train script occurred this error: """ ValueError: Expected input batch_size (21) to match target batch_size (2). """

How can I disable the BN layer? Thank you!

Hi! I want to know that how do you get the result. For Pascal VOC 2012，labeled ratio is 1/8，epochs is 34，Performing it for three times and getting the average of the last epoch i.e 33 epoch OR Performing it only once and getting the max result of 20 epoch to 33 epoch ?

thank you for release the code. we run the default config as your public in voc8.res50v3+.CPS, the best performance is 72.76, a little lower than the 73.20 in paper. have any other tricks? the default epoch is 34, the paper is 60? but your released log, 73.28 is also use 34 epoch.

In you few-supervise setting, which is in here which seems is labeled data but over lap with here.

Also, in you 1/8 ratio, labeled with unlabeled

in you 1/4 ratio, labeled with unlabeled,

in you 1/2 ratio, https://github.com/charlesCXK/TorchSemiSeg/blob/bab3a31ef7a5b59b932820b4d5d93a1b8bfc41ec/DATA/pascal_voc/subset_train_aug/pseudoseg_labeled_1-2.txt#L732 and https://github.com/charlesCXK/TorchSemiSeg/blob/bab3a31ef7a5b59b932820b4d5d93a1b8bfc41ec/DATA/pascal_voc/subset_train_aug/pseudoseg_unlabeled_1-2.txt#L321.

In each ratio, there is one overlap, and I suspect it should be a mis-operation.

Hi @charlesCXK, I was reading your paper and code, and I got a question concerning the numbers of epochs (max number of iters), did you choose it arbitrary ? Or do you take the fully supervised number of iters from deeplabv3+ as a starting point and adapt it to the various splits ?

I hope my question is clear :)

Hi,

Thank you for the sharing and excellent work. I have run the code for resnet50 but I fail to find the configs for resnet101, would you like to share them?

Besides, I notice there are some tiny changes in the resnet50 and ASPP. Is there any reason? and how to get the pre-trained model trained on ImageNet? (I mean using standard supervised learning or self-supervised learning?)

Besides, I try to download the city.zip and unzip it. But I fail to unzip it. So would you like to have a check?

By the way, thank you for your code, it is very helpful.

Best regards, Yuhang

First of all, please understand that I am asking you a similar question.

in your code ( https://github.com/charlesCXK/TorchSemiSeg/blob/main/exp.voc/voc8.res50v3%2B.CPS%2BCutMix/train.py ) and my previous question, ( https://github.com/charlesCXK/TorchSemiSeg/issues/60 )

I double-checked it several times, but DistributedSampler returns the same order of images in the same rank. ( DistributedSampler : shuffle=True , DataLoader : shuffle=False ) That is, since unsuperviced_train_loader_0 and unsuperviced_train_loader_1 use the same rank, the same image is called. (However, it is slightly different due to randomness of image augmentation)

Also, if you look at the shuffle condition code in the class DistributedSampler at the URL below, it is set to "g.manual_seed(self.seed + self.epoch)" . ( https://pytorch.org/docs/stable/_modules/torch/utils/data/distributed.html#DistributedSampler ) Therefore, each sampler is shuffled with "seed+epoch". -> each sampler is shuffled with seed+epoch, and your seed=0 (default) so the same image file is used for each epoch. (if same rank)

I think both unsuperviced_train_loader_0 and unsuperviced_train_loader_1 seeds should be different to mix different images using your code. ( i.e., unsuperviced_train_loader_0(train_dataset,seed=10+engine.local_rank) , unsuperviced_train_loader_0(train_dataset,seed=20+engine.local_rank) )

Like other recent studies, (i.e., PSMT cut-mix : https://github.com/yyliu01/PS-MT/blob/main/VocCode/train.py#L83 or U2PL cut-mix : https://github.com/Haochen-Wang409/U2PL/blob/b818f01c9c11cf6ebfb9fe4d679d2901eefa3f3c/u2pl/dataset/augmentation.py#L498 ) I think each should be mixed within the batch image.

I have checked several times, but unsuperviced_train_loader_0 and unsuperviced_train_loader_1 output the same image file within the same rank. Therefore, I think your code is mixed with the same image.

If I'm mistaken, I'd appreciate it if you let me know. Thank you!

Hi! a nice work! But I get some trouble when I try different label_ratio setting. The performance of the 1/4 is lower than the 1/8. Whether it should be smaller cps_weight with higher labeled_ratio? Any hint will be appreciate!

Hi, a nice work! However, I have a question about CPC（Cross probability consistency）. The results of the first and second calculations should be the same. Why should you keep the two?

Thanks.

Hi, there. Thanks for your paper, it's very inspiring! And I have a question to ask you, in Discussion you mentioned that the empirical results performs poorly in single-network pseudo supervision, for the network may tends to learn the wrong direction, but for multi-network, might play the regularizer for each other to prevent from going worst.

But now, for regularization consisitency, some paper proposed like Fixmatch, for the X is differently augmented( one is weak and one is strong) and hope the results keep consistent. And it worked.

What's the difference of the single-network mode and FixMatch? Just the augmentation way? Will the augmentation strength plays the main role in here?

Thanks a lot!

As I looked at your code, "model.eval()" is not used. (Instead, "torch.no_grad()" is used) ( ex. https://github.com/charlesCXK/TorchSemiSeg/blob/main/exp.voc/voc8.res50v3%2B.CPS%2BCutMix/train.py

line 209~219 )

However, I found that when each model generates a pseudo-label as a prediction, the accuracy of the pseudo-label is much higher when using "model.eval()" than when using "torch.no_grad()" .

Is this part okay? Or is the author's other intentions?

Thank you!