[Proposal] Compositional Configuration via Dataclasses (Hydra, etc.) #76
Assignees
Labels
enhancement
New feature or request
Comments
|
Following this proposal, I'll update with a working draft in PRs which we can look over and provide comments on. |
|
Thanks a lot for this! @romesco Looking forward to your proposal :) |
|
Adding reference to this issue IGE #127 on auto-completion when using Hydra. Seems pretty handy. |
5 tasks
|
Hi @romesco, I am wondering if there any updates on the plan with Hydra integration :) |
|
Thanks for the tag! I have something I'll push this weekend! Sorry about the tardiness, I had conference travel and then a deadline! |
Mayankm96
added a commit
that referenced
this issue
Jul 24, 2023
…sent (#76) # Description Previously, type annotation was always required to make the terms follow the order in which they are defined in the configclass. If this was not done, then the terms were getting sorted alphabetically which made it different from the expected behavior (user-defined order). On further inspection, turned out that in our wrappers for configclass, we were using `dir(cls)` to parse the class members, which sorts all the members of the class alphabetically. Changing it to `cls.__dict__` fixed this issue since in Python 3.7 onwards, dictionaries follow the user-defined ordering. Since this behavior changes the way config terms are parsed, the old configclass still exists inside the `omni.isaac.orbit.compat.utils.configclass` module so that people can still run policies trained with the old ordering. ## Type of change - Bug fix (non-breaking change which fixes an issue) - Breaking change (fix or feature that would cause existing functionality to not work as expected) ## Checklist - [x] I have run the [`pre-commit` checks](https://pre-commit.com/) with `./orbit.sh --format` - [x] I have made corresponding changes to the documentation - [x] My changes generate no new warnings - [x] I have added tests that prove my fix is effective or that my feature works - [x] I have updated the changelog and the corresponding version in the extension's `config/extension.toml` file
Mayankm96
added a commit
that referenced
this issue
Dec 22, 2023
…sent (#76) # Description Previously, type annotation was always required to make the terms follow the order in which they are defined in the configclass. If this was not done, then the terms were getting sorted alphabetically which made it different from the expected behavior (user-defined order). On further inspection, turned out that in our wrappers for configclass, we were using `dir(cls)` to parse the class members, which sorts all the members of the class alphabetically. Changing it to `cls.__dict__` fixed this issue since in Python 3.7 onwards, dictionaries follow the user-defined ordering. Since this behavior changes the way config terms are parsed, the old configclass still exists inside the `omni.isaac.orbit.compat.utils.configclass` module so that people can still run policies trained with the old ordering. ## Type of change - Bug fix (non-breaking change which fixes an issue) - Breaking change (fix or feature that would cause existing functionality to not work as expected) ## Checklist - [x] I have run the [`pre-commit` checks](https://pre-commit.com/) with `./orbit.sh --format` - [x] I have made corresponding changes to the documentation - [x] My changes generate no new warnings - [x] I have added tests that prove my fix is effective or that my feature works - [x] I have updated the changelog and the corresponding version in the extension's `config/extension.toml` file
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Proposal
Orbit has already made thoughtful design decisions that support really nice UX/UI paradigms for configuration management. Showing a few use-cases of how this might be done would be beneficial for the community as it can be difficult to decide "how to start". We would like to have one or two best-practices examples that fully leverage configuration via
dataclasses. Preferably, one which makes no commitment to configuration framework, and one which demonstrates a a few tweaks that can allow Hydra to be used at its most powerful capacity.Pitch
We'll provide an example that shows how to set up an environment(robot / scene), RL algorithm, and training script and configure all of this via dataclasses.
We'll then show how to extend this configuration paradigm via some key features of Hydra.
Alternatives
We can also look into what benefits other configuration frameworks have (one such being tyro), as long as they use
dataclassesas the principle data structure.Additional context
#57 and #67 is a related issue/PR pair, and it is a good start! We will build on this and provide more use-cases that leverage Hydra and are clean / easy to understand. See some of the previous discussion in #67 .
Checklist
The text was updated successfully, but these errors were encountered: