Quetion about searching reasults pbt_global.txt about pba HOT 7 CLOSED

@prob1995 wrote a great visualizer you can try. Unfortunately, the original visualization was done in google sheets / charts so there's no code.

from pba.

Hey, there's a namedtuple that defines each element in the list per line:

So the elements above correspond to each element in this tuple respectively.

from pba.

Should I use the augmentation strategy of the highest performing child-model to train my model? (For example, if the third child model have highest test accuracy, should I use pbt_policy_3.txt to train mine model ?)

from pba.

I am so sorry. I still can not understand the meaning of 'target_trial_name', 'clone_trial_name', 'target_trial_epochs', and 'clone_trial_epochs', 'old_config', 'new_config' .

from pba.

Yes, I'd recommend using the best performing child model to re-train.

The schedule params originate from https://github.com/ray-project/ray/blob/cdc9227f1b7c280465f76639ebc8c28ef7fc332e/python/ray/tune/schedulers/pbt.py#L253-L260

We're logging the trial ids, epoch/iteration numbers, and hyperparameter dict for the original worker and then one we're exploiting to.

from pba.

@arcelien could you share the codes to visualize the learned policy such as "Probability Hyperparameters Over Time" and "Magnitude Hyperparameters Over Time" in the github homepage?

from pba.

@snownus I wrote a .ipynb trying to reproduce the figure in the paper. Maybe you can use this.
https://github.com/prob1995/pba/blob/master/Policy_Visualization.ipynb

from pba.