v1.0.0 -- first stable release

We're pleased to announce the first stable release of imitation. Key improvements include:

• Gymnasium compatibility, which has superceded Gym
• Tuned hyperparameters and benchmark results for common algorithm-environment pairs (see release artifact attached).
• New algorithm (beta): SQIL
  For more information, see the changelog below.

What's Changed

• Updated Installation Instructions by @ernestum in #760
• Download experts from hf inside tutorials and docs by @jas-ho in #766
• Implementation of the SQIL algorithm by @RedTachyon in #744
• Additional examples of CLI usage by @EdoardoPona in #761
• Dependency fixes by @ernestum in #775
• Tune hyperparameters for kernel density estimation tutorial by @michalzajac-ml in #774
• Tune hyperparameters in tutorials for GAIL and AIRL by @michalzajac-ml in #772
• Introduce interactive policies to gather data from a user by @michalzajac-ml in #776
• Add an option to run SQIL with various off-policy algorithms by @michalzajac-ml in #778
• Complete PR #771 (Tune preference comparison example hyperparameters) by @lukasberglund in #782
• Add CLI for SQIL by @lukasberglund in #784
• Gymnasium Compatibility by @ernestum in #735
• Ensure MyST-NB raises an error when rendering a notebook fails. by @ernestum in #803
• Add a test timeout by @ernestum in #779
• Fix MacOS Pipeline: Include tests not in subdirectories by @AdamGleave in #797
• Remove MuJoCo dependency from SQIL notebook by @AdamGleave in #800
• Add partial support for dictionary observation spaces (bc, density) by @NixGD in #785
• Update gymnasium dependency and render_mode in gym.make by @taufeeque9 in #806
• Upgrade pytype by @ZiyueWang25 in #801
• Reduce training time and improve expert loading code in the tutorials by @ernestum in #810
• Add scripts and configs for hyperparameter tuning by @taufeeque9 in #675
• SQIL and PC performance check fixes by @ernestum in #811
• Running benchmarks by @ernestum in #812

New Contributors

• @jas-ho made their first contribution in #766
• @EdoardoPona made their first contribution in #761
• @michalzajac-ml made their first contribution in #774
• @lukasberglund made their first contribution in #782
• @NixGD made their first contribution in #785
• @ZiyueWang25 made their first contribution in #801

Full Changelog: v0.4.0...v1.0.0

v0.4.0

What's Changed

• Continuous Integration: Add support for Mac OS; remove dependency on MuJoCo
• Preference comparison: improved logging, support for active learning based on variance of ensemble.
• HuggingFace integration for model and dataset loading.
• Benchmarking: add results and example configs.
• Documentation: add notebook tutorials; other general improvements.
• General changes: migrate to pathlib; add more type hints to enable mypy as well as pytype.

Full Changelog: v0.3.1...v0.4.0

v0.3.1

What's Changed

Main changes:

• Added reward ensembles and conservative reward functions by @levmckinney in #460
• Dropping support for python 3.7 by @levmckinney in #505

Minor changes:

• Docstring and other fixes after #472 by @Rocamonde in #497
• Improve Windows CI by @AdamGleave in #495

Full Changelog: v0.3.0...v0.3.1

Major improvements

New features:

• New algorithm: Deep RL from Human Preferences (thanks to @ejnnr @norabelrose et al)
• Notebooks with examples (thanks to @ernestum)
• Serialized trajectories using NumPy arrays rather than pickles, ensuring stability across versions and saving space on disk (thanks to @norabelrose)
• Weights and Biases logging support (thanks to @yawen-d)

Improvements:

• Port MCE IRL from JAX to Torch, eliminating the JAX dependency. (thanks to @qxcv)
• Refactor RewardNet code to be independent from AIRL, and shared across algorithms. (thanks to @ejnnr)
• Add Windows support including continuous integration. (thanks to @taufeeque9)

First PyTorch release

compute_train_stats: Fix logits passed in as proba (#273)

Led to an error when I was training.

Final TF1 release

v0.1.1

Final TF1 release

Initial release

Prototype versions of AIRL, GAIL, BC, DAGGER.