How to run ml-agents on server?

Hello. I have been stuck with this for a while, and I need someone's help.

I manage to run a unity environment from my local (windows) machine. It's time now to move everything on the server to do proper ultra-fast computation. I build my scene using BuildTarget.StandaloneLinux64, and move it on the server. I am ssh'ing into a linux ubuntu server, and entering a docker container, using a Xming server, forwarding X11 on my machine. This works for demo apps (such as xclock or xeyes) but not with the unity environment:

Code (CSharp):

1. python code/examples/UnityMetaLearningTask.py
2. 2020-10-26 12:51:29 DEBUG [env_utils.py:30] The true file name is scene
3. 2020-10-26 12:51:29 DEBUG [env_utils.py:99] This is the launch string /work/./code/Unity-ML-Agents-Computer-Vision/Builds/Builds/MetaLearning_linux/N2_K4_Q2_sc64_dDatasetSupport/scene.x86_64
4. Found path: /work/./code/Unity-ML-Agents-Computer-Vision/Builds/Builds/MetaLearning_linux/N2_K4_Q2_sc64_dDatasetSupport/scene.x86_64
5. [[ AND AFTER SOME TIME... ]]
6.  "The Unity environment took too long to respond. Make sure that :\n"
7. mlagents_envs.exception.UnityTimeOutException: The Unity environment took too long to respond. Make sure that :
8.          The environment does not need user interaction to launch
9.          The Agents' Behavior Parameters > Behavior Type is set to "Default"
10.         The environment and the Python interface have compatible versions.
11.  

I have found this guide https://github.com/Unity-Technologies/ml-agents/blob/master/docs/Training-on-Amazon-Web-Service.md but I am afraid it's outdated and it won't lead me anywhere. So my question is: is there a better place to look for some guidance on that?
Thanks

 

Thanks! I have solved it few days ago thanks to that very Colab notebook. The important bit is to use the right xvfb version, as the one I had installed wasn't working properly. I am now able to run ml-agents *with visual observations* on a docker container in linux!
I have still trouble building a unity scene within a docker container, but that's a problem for another day..

 

Sure! Xvfb that worked: xvfb_1.18.4.
Not sure what was installed by default in my docker image (xvfb doesn't have a way to return the version :/

Anyway I have setup this repo to setup the Docker container with ML-Agents with visual observations, hopefully it can be useful
https://github.com/ValerioB88/ml-agents-visual-observations-docker

 

@ervteng_unity I managed to run a ml-agent scene on ubuntu on a docker container, on a server, with Visual Observations, with also exploiting the GPU for rendering the Unity scene itself. This is useful in case the scene involves a lot of graphic computation. Rendering stuff on the CPU may slow down training massively, as the network would constantly be waiting for new frames to arrive. In fact, my GPU was only active at 5% on average as couldn't get new frame fast enough.
In this way, my GPU is active at around 20%. Not great, but better. The network may still be waiting for new frames to arrive, and this may be due to socket speed or non-GPU related Unity stuff (as in my own script in the scene).

The Docker Image and relative instructions are here:
ValerioB88/ml-agents-visual-observations-docker-GPU: A sample Docker image for using ML-Agents with GPU support for Unity. (github.com)

 

I dont have Google access to this, can you please post a guide again onhow to set up training on a dedocated server?