Support heterogeneous dicts in infos

[jamartinh]

Infos is now a List[Dict]

Description

Allow for infos to be a list of heterogeneous dicts. Currently many environments use different keys for reset() and step(). Also wrappers are free to add arbitrary keys at arbitrary steps, for instance, RecordEspisodeStatistics add keys in the final step.

Hence, so save infos the users has to use almost always a custom step callback to regularize info data.

Fixes #246

Type of change

Please delete options that are not relevant.

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• This change requires a documentation update (perhaps)

Checklist:

• I have run the pre-commit checks with pre-commit run --all-files (see CONTRIBUTING.md instructions to set it up)
• I have run pytest -v and no errors are present.
• I have commented my code, particularly in hard-to-understand areas.
• I have made corresponding changes to the documentation.
• I solved any possible warnings that pytest -v has generated that are related to my code to the best of my knowledge.
• I have added tests that prove my fix is effective or that my feature works.
• New and existing unit tests pass locally with my changes.
• Backwards compatibility checks.
• New feature tests for backends: hdf5 and arrow.
• New parameter infos_format= None|"dict", "list" to add the optional functionality
• New parameter infos_format= None|"dict", "list" to add the optional functionality

[younik]

Thanks for the PR @jamartinh!

I didn't review it completely, just added a couple of comments. However, I do not agree on changing the infos to be a list of dict, as this approach is generally inefficient and inconsistent with what we do with Dict spaces.

To support heterogenous info, we should fill missing values with None/np.nans.
It should be possible to do it in EpisodeBuffer, I believe. This allows minimal changes in the rest of the codebase.

[jamartinh]

The feature is now optional, by specifying an optional parameter in data collector creation infos_format="list" otherwise when ìnfos_format=None (default) or infos_format="dict" all the functionality is exactly the same as default current version.

[jamartinh]

Infos can be saved as a dictionary of np.arrays or as a list of arbitrary dictionaries by
setting the infos_format parameter. Default is dictionary format infos_format = None or "dict":

from minari import DataCollector
import gymnasium as gym

env = gym.make('CartPole-v1')
env = DataCollector(env, record_infos=True, infos_format="list")

[younik]

Infos can be saved as a dictionary of np.arrays or as a list of arbitrary dictionaries by setting the infos_format parameter. Default is dictionary format infos_format = None or "dict":

from minari import DataCollector
import gymnasium as gym

env = gym.make('CartPole-v1')
env = DataCollector(env, record_infos=True, infos_format="list")

I don't like this for a couple reasons:

• Now the data format the user can expect when dealing with a dataset is uncertain. Even if it is just for the info, this weak the purpose of Minari which is, among others, to have a standardized format for offline RL data.
• Now every time we manipulate info, we basically need an if-else. Every branch is a different implementation to do the same thing, making the code less maintainable.