PBT tune Question

Ray Libraries (Data, Train, Tune, Serve) Ray Tune
1

Hi, everybody!
I have been trying the Population-Based Training (PBT) on the TD3 algorithm, and I have a minor question with respect to matching the hyperparameters. My question is " how do I equally set the ‘actor hidden nodes’ and ‘critic hidden nodes’ when I use the PBT on Tune of ray?".

For example, If I set as the below:
config[‘critic_hiddens’] = [tune.choice([[64, 64], [64], [32], [32, 32], [16, 16]])]
config[‘actor_hiddens’] = deepcopy(config[‘critic_hiddens’]),
tune sample each value for each config. I want to make these configs equal.

And the second question is “why does not each task run simultaneously?”

image
image1398×238 26.6 KB

Thank you!

2

Hey @Xim_Lee could you just use the same hyperparamter for both?

So inside your training function you can then just do

critic_hiddens = config["hiddens"]
actor_hiddens = config["hiddens"]

Let me know if this would work for you!

3

For the second question it’s probably because you don’t have enough resources to run more than 1 trial. Could you send the entire screenshot of the report table? That will show us how many resources you have.

4

Thank you @amogkam !
Your advices are helpful to me, and sorry for replying late.
I’ve been working on something else for a while.
I think I have enough resources, but I’ll upload my report table.