Release Notes

Next release plan

The next release should deliver the LunarLander environment benchamrk (almost done) and SAIL (IJCNN 2023).

v0.6.0

Support for Visual environments

This version adds support for visual environments (e.g., Atari) to all methods, the BaselineDataset, and the benchmark feature.

Known bugs:

  • For some reason, even though packaging builds the benchmark folder, it is not packaging it. So, for now, if users want to use the benchmark feature, they should install it from the source.
git clone https://github.com/NathanGavenski/IL-Datasets.git
cd IL-Datasets
pip install -e .

v0.4.0

Benchmarking

Now IL-Datasets has its own benchmarking! We are adding new methods and environments to the repository. For a full list of the methods and environments planned for release, please check the repository readme.md file.

Support for benchmark requirements

We split the imitation_datasets and benchmark modules requirements.

pip install il-datasets

will only install requirements regarding the imitation_datasets module. For using benchmark please use:

pip install "il-datasets[benchmark]"

Full Changelog: https://github.com/NathanGavenski/IL-Datasets/compare/0.3.0...0.4.0

v0.3.0

This version adds another point from the TODO list, Datasets!

Now, if you use the baseline_enjoy and baseline_collate functions, you can use the BaselineDataset. The datasets will load the generated numpy file and organize all entries to be (s_t, a_t, s_{t+1}), provide the average reward for all episodes and also allow for fewer episodes with the parameter n_episodes.

Alongside the dataset, I've implemented a HuggingFace solution as well as utility functions that allow users to upload their datasets to the HuggingFace website. There is already an example at: https://huggingface.co/datasets/NathanGavenski/CartPole-v1 In the future, these datasets will be used for benchmarking, but for now, it allows for storing outside drivers (such as Google's and Microsoft's)

This version also comes with some QoL improvements, such as pylint, and unit tests, so the code is more readable and also more stable.

Finally, with this release, I've implemented some metrics: performance, average episodic reward and accuracy.

Future release sneak peek

It is my plan that the future release will introduce benchmarking to IL-Datasets. With benchmarking, we will host a set of different datasets for common environments in the IL literature. This should help all researchers (including myself) to stop running different methods for each experiment.

Full Changelog: https://github.com/NathanGavenski/IL-Datasets/compare/0.2.0...0.3.0

V0.2.0

New Features

  • Added support for Gymnasium and Gym version 0.26.0.
  • Created template functions for enjoy and collate for a simple and one following the original dictionary from StableBaselines.

v0.1.0

FIx:

  • Sometimes, when the policy did not reach the goal and the enjoy function returned False, the Controller would not execute the enjoy function again

First release

First release for IL-Datasets. The missing features (such as documentation) list is in the README.md. If you have any issues with this release be sure to open an issue or contact me 😄