search

google-deepmind / acme

A library of reinforcement learning components and agents
Apache License 2.0
3.52k stars 426 forks source link

run_experiment allow different seeded/wrapped environments for train and eval #256

Closed Andrewzh112 closed 2 years ago

Andrewzh112 commented 2 years ago

current training loop seems to use only one environment object and usually eval env needs to seeded differently and wrapped differently (e.g. video wrapper on the eval env only). Is there a way to achieve this without editing the whole run_experiment function?

sinopalnikov commented 2 years ago

run_experiment is initially intended to cover only simple use cases. more features will be added to it going forward. for now you'd have to fork it and patch. or you can use make_distributed_experiment where ExperimentConfig.evaluator_factories are supported

Andrewzh112 commented 2 years ago

run_experiment is initially intended to cover only simple use cases. more features will be added to it going forward. for now you'd have to fork it and patch. or you can use make_distributed_experiment where ExperimentConfig.evaluator_factories are supported

alright got it thanks!!