How can we get the environment object from an algorithm object in general?
There are a couple of threads about this, with no robust answer.
- [rllib] Getting the env object from trainer · Issue #8700 · ray-project/ray · GitHub
- Getting the env object from trainer on ray 1.13.0
In the first thread the suggested solution is to use trainer.workers.local_worker().env where trainer = SomeAlgConfig(...).
This works for some algorithms, like DDPG (DDPGConfig), but not for others like PPO (PPOConfig).
Versions / Dependencies
Ray version 2.4.0
Python version 3.8
Reproduction script
Running this script I get:
ddpg env <TimeLimit<OrderEnforcing<PassiveEnvChecker<PendulumEnv<Pendulum-v1>>>>>
ppo env None
from ray.rllib.algorithms.ddpg import DDPGConfig
from ray.rllib.algorithms.ppo import PPOConfig
env_name = "Pendulum-v1"
# ---> DDPG <---
ddpg_config = DDPGConfig()
ddpg_trainer = ddpg_config.build(env_name)
print("ddpg env", ddpg_trainer.workers.local_worker().env)
# --------------
# ---> PPO <---
ppo_config = PPOConfig()
ppo_trainer = ppo_config.build(env_name)
print("ppo env", ppo_trainer.workers.local_worker().env)
# -------------
Issue Severity
High: It blocks me from completing my task.