Hi thank you for your reply! As for the hooks: when running the openmmlab training script the runner goes through some predefined functions, and by using hooks at certain 'mounting point' in the runner's workflow you can add some custom code to the training script. So I was thinking it might be possible to write a hook that initializes the clearml task at the beginning of the script and then calls execute_remotely(...) so the training is happening remotely. Anyhow, I'll look into the docker options of the clearml agent first. Thanks!

Hi @<1856869648880898048:profile|CourageousFly32> ,

Is it possible to run an entire docker container from a predefined docker image as a ClearML task (so that it's added to the queue like any other task)?

Yes. You can use the --docker flag in clearml-agent to make it run in agent mode - None

Or is it possible to train remotely using hooks in my MMDetection pipeline?

Can you elaborate?

Any other suggestion or possibilities? Best practices? I have very limited experience with Docker, ClearML and OpenMMLab. Thanks in advance!

I would suggest getting familiarized with Docker concepts and OpenMMLab first. ClearML adds capabilities on top and allows you to manipulate them better.