Best Practice Hyperparameter Tuning with long training duration

[Sbaetz]

First of all i tried finding similar questions but the only one that sounded similar was about a bug during hyperparameter tuning.

I am trying to train with yolov5x6 on a decently big dataset.
My training on said dataset takes about 2h per Epoch.
Since training time = tuning time (as far as i know), this can take quite a long time if i choose a high enough number of epochs to get a decent mAP50-95.

The big question is how can i efficiently hyperparameter tune in a timely manner and how does Yolov5 support this.

Should i only train with a fraction of my dataset (is there a option to only use samples) ?

Should i reduce my number of Epochs ?
If yes, then how do i determine my minimum epochs for the tuning ?
Is there a rule of thumb to have at least X% of mAP at that epoch or something like that ?
I already did some testruns with hyperparameter tuning so i could infere that point in my performance graphs.
(L Model for comparison: 20 Epochs: mAP50 = 0.95; mAP50-95 = 0.681)
Or do i simply use 1 Epoch to get as many results in my limited time ?

Can i hyperparameter tune on a yolov5 n or s model then adapt those hyperparameters to the x model ?
It may not translate well to the increase in model size but getting closer to great hyperparameters would be at least a good start.

Lastly i couldnt find Information about how i can train only specific Hyperparameters.
Is there an option during evolve ?
Can i create or modify any files to restrict changing certain parameters ?_
After excessively searching the forum i found a solution to "lock" certain hyperparameters:
Open train.py, go to the meta section and set mutation scale to 0 to not evolve certain hyper params.
#787 (comment)

Thank you in advance if you can take time to answer this.