Syllabus
Syllabus copied to clipboard
Integrate PLR
Prioritized Level Replay (PLR) is a simple-yet effective baseline for curriculum learning methods, and has been used in many prominent works ([AdA](https://arxiv.org/abs/2301.07608, MAESTRO, etc).
It would help adoption of Syllabus as well as curriculum research as a whole to have it implemented in a multiprocessing framework. Being able to benchmark against PLR with a single LOC change would be extremely helpful to my own research.
This will require:
-
[ ] A new version of the API that allows extra information to be passed to the curriculum.
-
[ ] The ability to manually pass information to the curriculum at any time, not just after task completion or during a step.
- It may be worth considering rolling the queues into one to simplify the API, or putting them all in a dict.
-
[ ] Ideally, a wrapper around the existing PLR implementation rather than a modification of it.
- Being able to import PLR rather than copy it would be huge for extending Syllabus to other curriculum learning methods. It would also be nice to have ready examples of how to do this.