Integrate MLflow Logging

[unniishere]

It will be good if you can integrate Mlflow logging like you guys did for wandb and tensorboard as wandb is paid.

[robmarkcole]

Plus 1

[lab176344]

PR is open #109

[isaacrob-roboflow]

what is the benefit of mlflow vs tensorboard?

[robmarkcole]

It is a server based solution, widely used by teams since it can be self hosted

[isaacrob-roboflow]

neat!

one thing I'm worried about is supporting an unlimited number of logging modules. would be convenient if we'd built on top of pytorch lightning and we'd get all these for free but 🤷 shipped on VERY short schedule.

@SkalskiP do you have thoughts on shipping too many logging frameworks? it does seem like there's a good argument for this one

[isaacrob-roboflow]

we should make sure to keep these learnings in mind for the next model we ship

[robmarkcole]

I've had conversations with lightning about logger support - they basically support log metric and log_dict as standard. It gets logger specific for artefact logging. Overall +1 for using lightning for the training loop - this was a factor that held me back from embracing ultralytics

[isaacrob-roboflow]

yeah I am a really big fan of pytorch lightning, just didn't make sense to refactor for this project and probably too late now

[lab176344]

@isaacrob-roboflow we host mlflow servers in gcp, has almost the same functionalities as W&B. Ultralytics has support for Mlflow too. +1 for lighting too =)

[isaacrob-roboflow]

@SkalskiP I'm feeling down to have mlflow support in rf-detr repo

[ggand0]

+1

[Akhp888]

+1

[robmarkcole]

@isaacrob-roboflow @SkalskiP the PR is ready, could really use this functionality :-)

[dkrutika]

+1

[SkalskiP]

Hi all! I promise to take a look at the PR this week.