Skip to content

v2.6.0: Refactored hyperparameter optimization

Latest
Latest
Compare
Choose a tag to compare
@araffin araffin released this 24 Mar 15:07
· 2 commits to master since this release
v2.6.0
325ef5d
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Breaking Changes

  • Upgraded to SB3 >= 2.6.0
  • Refactored hyperparameter optimization. The Optuna Journal storage backend is now supported (recommended default) and you can easily load tuned hyperparameter via the new --trial-id argument of train.py.

For example, optimize using the journal storage:

python train.py --algo ppo --env Pendulum-v1 -n 40000 --study-name demo --storage logs/demo.log --sampler tpe --n-evaluations 2 --optimize --no-optim-plots

Visualize live using optuna-dashboard

optuna-dashboard logs/demo.log

Load hyperparameters from trial number 21 and train an agent with it:

python train.py --algo ppo --env Pendulum-v1 --study-name demo --storage logs/demo.log --trial-id 21

New Features

  • Save the exact command line used to launch a training
  • Added support for special vectorized env (e.g. Brax, IsaacSim) by allowing to override the VecEnv class use to instantiate the env in the ExperimentManager
  • Allow to disable auto-logging by passing --log-interval -2 (useful when logging things manually)
  • Added Gymnasium v1.1 support

Bug fixes

  • Fixed use of old HF api in get_hf_trained_models()

Other

  • scripts/parse_study.py is now deprecated because of the new hyperparameter optimization scripts

Full Changelog: v2.5.0...v2.6.0

Assets 2
Loading