record_video.py not working on PyBullet envs [bug]

[dlr-rm]

record_video.py not working on PyBullet envs [bug]

[araffin]

you have two solutions : try with MKL_THREADING_LAYER=GNU (see https://github.com/DLR-RM/stable-baselines3/blob/master/.gitlab-ci.yml) or change the default start method to fork (cf doc).

[mcres]

I don't follow, could you please be more specific? Where should I make those specific changes?

Also, would be fixing this doable? If not, handling the error and display a message to how to manually fix it sounds like a good option to me.

[araffin]

I don't follow, could you please be more specific? Where should I make those specific changes?

MKL_THREADING_LAYER=GNU is an environment variable, so it works like that:
MKL_THREADING_LAYER=GNU python train.py ...

the start method can be specified here:

https://github.com/DLR-RM/rl-baselines3-zoo/blob/master/utils/exp_manager.py#L91

Also, would be fixing this doable? If not, handling the error and display a message to how to manually fix it sounds like a good option to me.

unfortunately, it may come from python and on some linux computers only, so I'm not sure how easy it is to catch...

[mcres]

Thanks! I tried both methods but unfortunately, I keep getting the same error :/
What I don't understand is why it's trying to create several pybullet envs to create a single video. Isn't one env enough for that?

unfortunately, it may come from python and on some linux computers only, so I'm not sure how easy it is to catch...

I guessed it wasn't easy, at least this issue can be useful for the moment

[araffin]

So, the fix is here: #77

Different things:

• as mentioned in the doc (https://stable-baselines3.readthedocs.io/en/master/guide/vec_envs.html) the thread safe method need to be protected with the __main__ (I wrote the code when we were still using fork as default)
• the start method was not at the right place, for testing, it is defined here: https://github.com/DLR-RM/rl-baselines3-zoo/blob/master/utils/utils.py#L203
• we are not using multiple envs, but just one subproc env for rendering PyBullet envs (this is a hack), you can also pass --no-render if you just need the video without direct visual feedback (faster).