training problems

[wenzhi-zhang]

when l train model in real world dataset,l meet some error and the final result is differ from the paper. this is the error during the training and this is the training result

[jkulhanek]

Thanks for sharing. I will take a look at it as soon as I find some time. I guess the error is in unreal auxiliary tasks. I refactored the deeprl-pytorch package since the release of the paper and I might have caused some errors.

[GongYanfu]

Hello, are you still using the author's model? Have you used the author's code on a real robot? Can it run smoothly on a robot? Can it run successfully on the server? Recently I‘m trying to use the pre-trained model by author on my turtlrbot2, but I cannt run it smoothly. Fisrtly there some errors in ros workspace and then I resolved it, but now there are some errors in python folder such as agent.py, model.py. So I want to know if you run it successfully? I'd be grateful if you can answer my question. Thanks

[jkulhanek]

The source code published in the repository is not the original one I used for the experiments. When I was cleaning it perhaps I added a bug causing the auxiliary tasks not to work properly. I will try to compare the original source code side by side with the new one and locate the bug. If anyone wants it I can also share with you the original source code (but it is a bit messy).

[jkulhanek]

Please post any errors you may encounter (open a new issue for the error) so I can look at them.

[GongYanfu]

存储库中发布的源代码不是我用于实验的原始代码。当我清理它时，也许我添加了一个导致辅助任务无法正常工作的错误。我将尝试将原始源代码与新源代码并排比较并定位错误。如果有人想要它，我也可以与您分享原始源代码（但有点乱）。

can you share the original source code here? if you can , upload it to github, thank you.

[jkulhanek]

The source code is here: https://data.ciirc.cvut.cz/public/projects/2021RealWorldNavigation/checkpoints/old-rw-nav.zip The corresponding version of deep-rl-pytorch should be 0.3.0!

[jkulhanek]

I guess the deep-rl-pytorch version is even older. I will try to find it.

[jkulhanek]

Ok, please try 0.2.7

[GongYanfu]

Ok, please try 0.2.7

There is no deep-rl-pytorch V0.2.7, I found the earliest version is 0.3.1

[jkulhanek]

What is the output of pip install deep-rl==0.2.7?

[GongYanfu]

The source code published in the repository is not the original one I used for the experiments. When I was cleaning it perhaps I added a bug causing the auxiliary tasks not to work properly. I will try to compare the original source code side by side with the new one and locate the bug. If anyone wants it I can also share with you the original source code (but it is a bit messy).

I think the error occurs in the code of your deep-rl-pytorch repo, not the code of robot-visual-navigation?

[jkulhanek]

I agree, therefore use the older deep-rl-pytorch version.

[GongYanfu]

I used all of the other version of deep-rl-pytorch. if I use v0.3.6, there are same problem. If I use other older versions, there will be new problem.

[jkulhanek]

You should try 0.2.7.

[GongYanfu]

I dont use it in the old-rvn, I'm trying to use deep-rl in the newest rvn. I cannt use v0.2.7, its too low.

[jkulhanek]

Ok. I know about the bug and will try to find the cause of the problem with the current version when I have some time. Meanwhile please use the older version of this repo with the older version of deep-rl-pytorch to reproduce the paper's results.

[GongYanfu]

It seems that the training results are still somewhat abnormal by running the code of old-version repo, where the red represents the experiment ”turtlebot“, and the green represents the "turtlebot-a2c"

[jkulhanek]

I have merged the old source code to the current version (PR #15). The issue should be resolved now. I tried training the turtlebot experiments and it is working now.

If you still encounter problems with training please open a new issue.