Find a better way to convert notebooks to tutorial pages about flygym HOT 9 OPEN

@sibocw, I'll look into colab issue. I wanted to take a deeper look at this stuff but I didn't have a good cause :)

from flygym.

Nope, I meant that I wanted to get more familiar with Jupiter/Colab in general, I'm a C++ developer in a normal life :)

from flygym.

Maybe it makes sense to make the tutorials available in colab, just like MuJoCo does? It lowers the entry barrier significantly.

https://mujoco.readthedocs.io/en/latest/python.html#tutorial-notebook

from flygym.

We tried Colab last year when using a primitive version of this package for a course. We moved away from it because

1. When using dm_control, running multiple simulations (even sequentially) caused the Colab runtime to crash due to some threading issue with the EGL renderer. We haven't encountered this problem on any other environment.
2. We wanted the web page to be fully viewable (including videos) without the user having to run the code.
3. It was very slow to run things on Colab that month for some reason (GPUs were scarce maybe?). I don't remember what happened exactly.

Regarding 1, I just tried it again and encountered the same issue. This notebook crashes the Colab runtime at the last cell. The following is the error stack:

Exception ignored in: <function MjrContext.__del__ at 0x7a65e2238e50>
Traceback (most recent call last):
  File "/usr/local/lib/python3.10/dist-packages/dm_control/mujoco/wrapper/core.py", line 638, in __del__
  File "/usr/local/lib/python3.10/dist-packages/dm_control/mujoco/wrapper/core.py", line 629, in free
  File "/usr/lib/python3.10/contextlib.py", line 135, in __enter__
  File "/usr/local/lib/python3.10/dist-packages/dm_control/_render/base.py", line 135, in make_current
RuntimeError: Cannot make context <dm_control._render.pyopengl.egl_renderer.EGLContext object at 0x7a65da5c9f00> current on thread <Thread(_colab_inspector_thread, started daemon 134578698184256)>: this context is already current on another thread <_MainThread(MainThread, started 134579582074880)>.

This only happens on the second simulation of the same notebook and only on Colab. It seem related to a "known issue with dm control" according to this issue, but the solution proposed there didn't seem immediately applicable to Colab.

So in the end we decided that it was not worth the effort, as we had a tight timeline for the course. Getting this to work could also be pretty nice.

from flygym.

@stepanove Thank you! Did you encounter this issue before? Was it also with Golab?

from flygym.

@sibocw, it looks like this exception doesn't really break anything (it says it's ignored). I had it few times but the video was generated anyway.
Here is my test notebok. It renders raw frames to ensure it doesn't show the video from previous run.

from flygym.

IMHO, an interactive tutorial is way better than just a web page. If I'm not involved into the process I just scroll it paying attention to videos only :)

btw, the size of the notebook with videos is just 5MB.

from flygym.

Hi @stepanove

it looks like this exception doesn't really break anything (it says it's ignored). I had it few times but the video was generated anyway.

Interesting! For me it certainly crashed the simulation because the whole Jupyter runtime crashed. I will try to look into this more.

IMHO, an interactive tutorial is way better than just a web page. If I'm not involved into the process I just scroll it paying attention to videos only :)

I totally agree. If Colab works we can at least provide a link to play with the tutorials on Colab (in any case, the notebook can be opened on Colab simply with the URL https://colab.research.google.com/github/xxx/xxx/xxx.ipynb. Our notebooks are already on Github so this should be easy).

btw, the size of the notebook with videos is just 5MB.

My concern is that the videos might be updated and I don't want another 5MB of data to be tracked by Git every time the notebook changes. This is especially important because the package might be installed/cloned in ephemeral environments (Github Actions, or even Colab notebooks where the package needs to be reinstalled for every runtime).

from flygym.

My concern is that the videos might be updated and I don't want another 5MB of data to be tracked by Git every time the notebook changes. This is especially important because the package might be installed/cloned in ephemeral environments (Github Actions, or even Colab notebooks where the package needs to be reinstalled for every runtime).

Yep, I also think that storing videos in git (at least in the main package repo) is not good.
I heard github was going to add GPU enabled CI runners, then it can be used to execute notebooks and deploy the results somewhere.

from flygym.