Hydra exploration by ernestum · Pull Request #703 · HumanCompatibleAI/imitation

First draft with somehwat working airl configured using Hydra.

ernestum

committed

Use recursive calls and introduce random number generator config.

ernestum

committed

Split up AILR config in AIRL creation and AIRL run config. Also move most of the interpolation magic into the main script and rename some util modules to make clear what they do.

ernestum

committed

Make inline imports to speed up shell completion.

ernestum

committed

Add type annotations, fix typing issues, add comments.

ernestum

committed

Remove env_test.py

ernestum

committed

Add code for policy evaluation.

ernestum

committed

Move airl configuration to it's own file and restructure the main run file

ernestum

committed

Add checkpoint saving.

ernestum

committed

Use the hydra chdir feature to store logs in the output directory.

ernestum

committed

Remove defaults parameter from `register_configs`, introduce a air_run.yaml and assemble the airl run config when registering in the config store.

ernestum

committed

Define cartpole and pendulum envs as structured configs.

ernestum

committed

Ensure PPO on disk does not inherit from PPO and loads from an absolute path.

ernestum

committed

Remove low default number of steps for generated trajectories and move that to the config file.

ernestum

committed

Introduce default_environment to the register_configs functions, remove the global demonstrations field and rename the global venv field to environment.

ernestum

committed

Move registering the expert policy as a sub-call to registering the trajectories.

ernestum

committed

Update the airl_sweep.yaml

ernestum

committed

Update the airl_sweep.yaml

ernestum

committed

Add airl_optuna.yaml

ernestum

committed

Formatting, typing and documentation fixes. Also the implicit seed dependency was pulled out of the utils and made explicit in airl.py

ernestum

committed

Swtich from Hydra call to hydra instantiate.

ernestum

committed

Add type ignore reason.

ernestum

committed

Dont' allow variable horizon for AIRL by default.

ernestum

committed

Simplify the class configurations using enums.

ernestum

committed

Fix bug in type ignore reason.

ernestum

committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Hydra exploration#703

Hydra exploration#703
ernestum wants to merge 25 commits intomasterfrom
hydra_exploration

Commits on Jun 1, 2023

First draft with somehwat working airl configured using Hydra.

Use recursive calls and introduce random number generator config.

Split up AILR config in AIRL creation and AIRL run config. Also move most of the interpolation magic into the main script and rename some util modules to make clear what they do.

Make inline imports to speed up shell completion.

Add type annotations, fix typing issues, add comments.

Remove env_test.py

Add code for policy evaluation.

Move airl configuration to it's own file and restructure the main run file

Add checkpoint saving.

Use the hydra chdir feature to store logs in the output directory.

Remove defaults parameter from `register_configs`, introduce a air_run.yaml and assemble the airl run config when registering in the config store.

Define cartpole and pendulum envs as structured configs.

Ensure PPO on disk does not inherit from PPO and loads from an absolute path.

Remove low default number of steps for generated trajectories and move that to the config file.

Introduce default_environment to the register_configs functions, remove the global demonstrations field and rename the global venv field to environment.

Move registering the expert policy as a sub-call to registering the trajectories.

Update the airl_sweep.yaml

Update the airl_sweep.yaml

Add airl_optuna.yaml

Formatting, typing and documentation fixes. Also the implicit seed dependency was pulled out of the utils and made explicit in airl.py

Swtich from Hydra call to hydra instantiate.

Add type ignore reason.

Dont' allow variable horizon for AIRL by default.

Simplify the class configurations using enums.

Fix bug in type ignore reason.