lslrh / cpsl Goto Github PK

Hi, nice work.
I wonder if you can share the models adapted from SYNTHIA to Cityscapes?
I would appreciate it if you could share them with us.

请问这个模型文件是用 源域+源域标签 用DeepLab有监督训练得到的吗

Could you please also share the training script to reproduce the results?
thanks a lot.

您好，从您定义的公式（9）来看，每个类别采样的像素值是根据伪标签中各个类别占比确定的，因此M中的各个类别的像素个数还是不平衡的。请问这种采样方式为什么能解决长尾为题？

你好，您的工作让人感兴趣。在复现您工作的时候遇到了一点问题，需要您的帮助。
在进行stage1训练时， train.py ，Line43执行时缺乏prototypes_on_cityscapes_from_deeplabv2.pth文件。
我运行calc_prototype.py 生成的文件是一个.zip的压缩文件，这是为什么呢？
请问我应该怎么操作才能得到prototypes_on_cityscapes_from_deeplabv2.pth这个文件呢？

I want to train the model from the very beginning. Would you please provide the script for the warm-up phase?

@lslrh Thank you for opening source code.
Why you chose the weight parameter of the best model of Stage1 to generate soft pseudo label of Stage3 rather than use the weight parameter of the best model of Stage2?

• Stage3 Generate soft pseudo label.
  python generate_pseudo_label.py --name gta2citylabv2_stage2 --flip
  --resume_path ./logs/gta2citylabv2_stage1Denoise/from_gta5_to_cityscapes_on_deeplabv2_best_model.pkl --no_droplast
  --bn_clr --student_init simclr

Hello. First, I want to thank you for this code. It helped me a lot.

I’ve been reproducing the results on my own, using four A5000 gpus.

At phase one, the performance was just the same which is 55.6 mIoU as you wrote in the readme file. However, I think there are some issues at phase two since the performance has dropped to 58.2, 1 mIoU lower compared to your work. I’m not clear whether the cause of the issue is me or not.

So if there is someone who tried reproducing this work and been through this issue, please help me.

I will be appreciated so much. Thank you.

Hello, thanks for sharing a code of your amazing work.
Recently, I rerun your code for stage 1, but the result (52.9 mIoU) is slightly lower than reported (55.7 mIou).

I follow the Readme file and do not change the default setting.
Our best result is as below:
2022-08-27 15:37:37,300 INFO Overall Acc: : 0.8905403760076205
2022-08-27 15:37:37,300 INFO Mean Acc : : 0.6646033595962516
2022-08-27 15:37:37,301 INFO FreqW Acc : : 0.8140811865214824
2022-08-27 15:37:37,301 INFO Mean IoU : : 0.5294283467093287
2022-08-27 15:37:37,301 INFO 0: 0.9056666268706702
2022-08-27 15:37:37,301 INFO 1: 0.5498352885675489
2022-08-27 15:37:37,301 INFO 2: 0.8142533757567021
2022-08-27 15:37:37,301 INFO 3: 0.3985532715974026
2022-08-27 15:37:37,301 INFO 4: 0.2720654507937554
2022-08-27 15:37:37,301 INFO 5: 0.45106496224026094
2022-08-27 15:37:37,301 INFO 6: 0.5157289767810124
2022-08-27 15:37:37,301 INFO 7: 0.394099157886852
2022-08-27 15:37:37,301 INFO 8: 0.8491513259166265
2022-08-27 15:37:37,301 INFO 9: 0.43677405299954
2022-08-27 15:37:37,301 INFO 10: 0.7520329908008514
2022-08-27 15:37:37,301 INFO 11: 0.6789538247878648
2022-08-27 15:37:37,301 INFO 12: 0.20250515414401246
2022-08-27 15:37:37,301 INFO 13: 0.8740814618033623
2022-08-27 15:37:37,301 INFO 14: 0.4327794107544921
2022-08-27 15:37:37,301 INFO 15: 0.5791000163060087
2022-08-27 15:37:37,301 INFO 16: 0.10156630850796927
2022-08-27 15:37:37,302 INFO 17: 0.38720734417824565
2022-08-27 15:37:37,302 INFO 18: 0.46371958678406566

How can I reproduce your work in GTA5->Cityscapes scenario?

I also attached our log file below.
run_2022_08_26_05_49_33.log

thank you.

Hi! In your paper, the features of the target domain images are clustered and assigned the pseudo labels Psl by the Cluster Assignment. But I haven't found the corresponding code of this part and could you please give me a hint? Thanks a lot!

Hello, I'm very interested with your amazing work.

I rerun your code, but the ‘cutmix’ augmentation in the stage1 phase does not seem to work properly.
Below is the visualization result of input image (target_x), prediction (target_out[‘out’]), and pseudo gt (threshold_arg) in order, all of which are applied to ‘cutmix’ augmentation.
In my opinion, it seems to be a problem caused by the difference between the label size and the input size in the process of augmentation of target PGT.
Let me know if I’ve misunderstood anything.

Thank you very much.

One 3090GPU have 24GB. ProDA and your code write "to reproduce the performance, you need 4 GPUs with no less than 16G memory."
So I want to know if it can use two GPU reproduce the performance.