codestar12/experimental

mcontrib

mcontrib is MosaicML's contrib repository for algorithms. Have a cool new efficiency method to experiment with, or want to get your published work used by industry? Contribute new algorithms here, and we will work with you to benchmark them and upstream them into Composer!

For an in-depth explanation of how to write custom speed-up methods, see our notebook here.

Adding algorithms

To add an algorithm to mcontrib, create a folder mcontrib/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!
• [Optionally] README.md briefly describing your algorithm.

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 mcontrib

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

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

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

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

import mcontrib
from composer.trainer import TrainerHparams

mcontrib.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