Is it possible to use tune.with_resources() with str as alternative to trainable?

The tune.Tuner() allows using string, e.g. ‘PPO’ as first argument. However, Tuner doesn’t provide any way to input ScalingConfig. On the other hand, tune.with_resources() can set resource limit on a trainable but won’t take a string as input. This raised an interesting question: is there a way to set resource limit on tune.Tuner() when using PPO in rllib?

Given tune.Tuner() accepts both trainable and str, I thought internally it is converting a str to a trainable. However, looking at the source code I didn’t see where it is doing the conversion

To set PPO trainable resources, you probably should leverage PPOConfig class. See Algorithms — Ray 2.2.0

cc @kourosh

1 Like