experimental

experimental is a repository of algorithms that are either third-party contributions, or not sufficiently mature to contribute into Composer.

Adding algorithms

To add an algorithm to experimental, create a folder experimental/algorithms/your_algo_name, with the following files:

• __init__.py that imports your algorithm class
• metadata.json with some metadata for your algorithm.
• *.py with your code!

The metadata.json should have the following fields:

{
    "name": "My Example Algorithm",
    "class_name": "ExampleAlgorithm",
    "tldr": "Just an Example",
    "attribution": "(Example et al, 2022)",
    "link": ""
}

Where the "class_name" field should be importable from your_algo_name folder. The other fields are optional.

For an example, see ExampleAlgorithm.

Using experimental

To use experimental in your code, simply import the library and use with your trainer:

from experimental.algorithms.example_algorithm import ExampleAlgorithm
from composer import Trainer

trainer = Trainer(
    algorithms=[ExampleAlgorithm()],
    ...,
)

To use experimental with YAHP and our YAML config files, in your entrypoint code, call register_all_algorithms(), after which the algorithms will be accessible through YAHP, our config management library.

import experimental
from composer.trainer import TrainerHparams

experimental.register_all_algorithms()
trainer = TrainerHparams.create(f="my_yaml_file.yaml")
trainer.fit()

The key in your YAML file is the folder name of the algorithm:

algorithms:
  - example_algorithm:
    alpha: 0.1