Reproducibility of the codes

[ChineseYjh]

Hello~I'm interested in your work. However, your codes cannot reproduce the results of Tab.I in your paper, which is the mIoU results with the FPN on Cityscapes val set, even though we tune the hyperparameters the same as your paper describes. The best we achieve is 77.07%, far from your 78.02%.

Could you offer more details to reproduce the corresponding results, especially maybe the arguments in eval.py and config.yaml?

[lpc-97]

你好，可以要一份你运行的代码么，卡在一个步骤上了。感谢：liangpengchen@shu.edu.cn

[lpc-97]

你好，可以要一份你运行的代码呢，卡在一个步骤上了。谢谢：liangpengchen@shu.edu.cn \

你好~我对你的作品很感兴趣。但是，您的代码无法重现您论文中**Tab.I的结果，即使用 FPN 在 Cityscapes 验证集上的 mIoU 结果，即使我们按照您的论文描述的方式调整超参数。我们达到的最好成绩是77.07%**，与您的 78.02% 相去甚远。

您能否提供更多详细信息以重现相应的结果，尤其是_eval.py_和_config.yaml_中的参数？

你好，可以要一份你运行的代码呢，卡在一个步骤上了。谢谢：liangpengchen@shu.edu.cn

[monoelh]

Hello~I'm interested in your work. However, your codes cannot reproduce the results of Tab.I in your paper, which is the mIoU results with the FPN on Cityscapes val set, even though we tune the hyperparameters the same as your paper describes. The best we achieve is 77.07%, far from your 78.02%.

Could you offer more details to reproduce the corresponding results, especially maybe the arguments in eval.py and config.yaml?

Hi, I'm also curious about some details. Which parameters did you adapt to reach 77.07% mIoU? The config.yaml looks like a baseline to me. Especially since it uses _ultrape and not _learnpe. If you didn't change this parameter, then 77.07% looks more in line with the results of Table 4.
However, the parameters local and unfold remain a mystery to me.