Skip to content

CONTRIBUTING.md

Latest commit

 

History

History
49 lines (41 loc) · 2.76 KB

CONTRIBUTING.md

File metadata and controls

49 lines (41 loc) · 2.76 KB

Contributing to metalearn

Adding New Metafeatures

metalearn uses a caching mechanism to cache expensive computations that may need to be used again within the package by another function. Both resources (e.g. the dataset itself or a preprocessed version of it) and metafeatures (e.g. entropy, number of features) are cached by the system.

When adding a new metafeature to the package, the function that computes that metafeature needs to be included in the Metafeatures class definition in the Metafeatures._mfs_info attribute in ./metalearn/metafeatures/metafeatures.py. Before the function can be included though, it needs to be decorated with metadata by being passed through the MetafeatureComputer constructor (see example below). This allows the metafeatures returned by the function to be used intelligently by the package.

Follow the example below to know how to write and register new metafeature(s). Note that a metafeature-computing function (e.g. get_dataset_stats as seen below) can compute and return more than one meta-feature.

# Import needed utilities
from metalearn.metafeatures.base import MetafeatureComputer
from metalearn.metafeatures.constants import ProblemType, MetafeatureGroup

# Declare the function that computes the metafeatures.
def get_dataset_stats(X, column_types):
    # Calculate metafeatures.
    number_of_instances = X.shape[0]
    number_of_features = X.shape[1]
    # Return a tuple (here it's two metafeatures).
    return (number_of_instances, number_of_features)

# Decorate the metafeature-computing function with data
# the package will use.
get_dataset_stats = MetafeatureComputer(
    # Pass the function into the `MetafeatureComputer`
    # decorator.
    computer=get_dataset_stats,
    # Give each metafeature returned by the function a
    # name for the cache to use (order here must match the
    # order they are returned in by `computer`).
    returns=[
        "NumberOfInstances",
        "NumberOfFeatures"
    ],
    # Associate a problem type with the new metafeatures.
    problem_type=ProblemType.ANY,
    # Associate one or more metafeature groups.
    groups=[MetafeatureGroup.SIMPLE],
    # Specify which values to pass to the function
    # when calling it to compute the metafeatures.
    # Here we are passing the cached resource called
    # "X_raw" as the value for this function's "X" argument.
    argmap={ "X": "X_raw" }
)

By convention, all the decorated metafeature-computing functions in a module are aggregated at the bottom of the module into a call to build_resources_info, which is then imported by ./metalearn/metafeatures/metafeatures.py and added to the _mfs_info attribute of that module's Metafeatures class.