Use-Cases of APEBenchยค
I want to:
- Benchmark my neural architecture:
- Is your architecture implemented in Equinox? Then you can directly register it with APEBench and use the entire pipeline.
- Your architecture is not implemented in Equinox. Then you can not use the full pipeline of APEBench (no differentiable physics, for example). But you can still scrape datasets for purely data-driven training in different ecosystems and you can evaluate your architectures (TODO add link) by rolling them out in your ecosystem of choice and casting the results back to JAX arrays and using APEBench's metric system.
- Create datasets Consider using the scraping API.
- Conduct studies Learn about the typical workflow of APEBench and consider using the command line interface (CLI).
- Extend APEBench:
- With new architectures. If this is just for your experiments, then you
can register them in the global
architecture_dict
. If you want to expand the built-in architectures, consider opening a pull request to PDEquinox. - With new scenarios. You can subclass from
BaseScenario
(or other child classes) and register them in thescenario_dict
. - New activation functions, initial condition distributions, optimization and learning rate scheduling setups, etc. can be registered in their respective components dictionary.
- With new architectures. If this is just for your experiments, then you
can register them in the global
- Reproduce the experiments of the APEBench paper: All experimental scripts (and also the results) are available here and the results of the ablational experiments are here.