Reproduce the segmentation result

[shiyutang]

I have tried to rerun the segmentation part of the code in python 2.7
and I was unable to reproduce the result in the paper.
I have noticed several differences.

1. I have rewritten the label map part in the GTA dataset since GTA needs to map to Cityscape and the original code didn't do it.
2. I was using torch 1.3.1
3. I update the loss to NLLLoss rather than NLLLoss2d
4. I fixed a few grammar bugs which probably happened because of the change of time in the code
   here is my result:

I would love to see if anyone was able to reproduce the segmentation result, and I would be really appreciated if you could give me that piece of code such that I could see where went wrong.

THANKS, A LOT!

[shiyutang]

the training process with different learning rate setting is shown below:

[Lufei-github]

看你有跑通这篇代码，所以想请教你个问题！我是python2.7,pytorch1.0跑的segmentation的代码！

看到有报这个错误，RuntimeError: cuda runtime error (59) : device-side assert triggered at ...

网上说是计算损失的时候label不是从0开始的，pytorch必须是（0，n-1）的label才不会报错，然后这篇论文在设置label的时候，类别个数设置为20，有一个类别作者说是背景。但是我现在想，我的这个报错会不会与作者设置的20个label有关！
请问你在跑这篇代码的时候有遇到这个问题吗？

或者你知道这是哪里出错了吗？多有打扰，期待你的答复！！！

[happyzhouch]

看你有跑通这篇代码，所以想请教你个问题！我是python2.7,pytorch1.0跑的segmentation的代码！

看到有报这个错误，RuntimeError: cuda runtime error (59) : device-side assert triggered at ...

网上说是计算损失的时候label不是从0开始的，pytorch必须是（0，n-1）的label才不会报错，然后这篇论文在设置label的时候，类别个数设置为20，有一个类别作者说是背景。但是我现在想，我的这个报错会不会与作者设置的20个label有关！
请问你在跑这篇代码的时候有遇到这个问题吗？

或者你知道这是哪里出错了吗？多有打扰，期待你的答复！！！

你好，我对作者关于这两个数据"synthia", "city16"设置的16个类别这里有疑问， 0-15是15个实际的类别，255是背景，那在代码里args.n_class就等于16， 那么这句话： weight[args.n_class - 1] = 0 # Ignore background loss， 其实是将标签为15 的类别的weight设为0了，而不是将背景weight设为0的啊？