Params

Contains the base parameter class ExperimentParameters, a dataclass meant to represent a configuration for an experiment run.

Classes:

ExperimentParameters(name, hash, overwrite, …)

Base parameter class, for handling naming and hashing.

class curifactory.params.ExperimentParameters(name: str = 'UNNAMED', hash: Optional[str] = None, overwrite: bool = False, hash_representations: dict = <factory>)

Base parameter class, for handling naming and hashing.

In any given repo, this class should be extended to contain any needed local configuration.

Note

Extending with a @dataclass is recommended to make it syntactically easier to read and define.

Example

from dataclasses import dataclass
from curifactory import ExperimentParameters

@dataclass
class Params(ExperimentParameters):
    some_parameter: int = 0
    # ...

Attributes:

hash	Curifactory automatically fills this, but it can be overriden if you need to use very specific cache naming.
hash_representations	Dictionary of parameter names in the dataclass where you can provide functions that return a unique/consistent representation for each parameter to use as the value for hashing.
name	Parameter set name.
overwrite	Whether to overwrite pre-cached values.

Methods:

params_hash([dry])

Convenience function to see the hash of these parameters that curifactory is computing, or debug them with dry=True.

hash: str = None

Curifactory automatically fills this, but it can be overriden if you need to use very specific cache naming. (Should not normally be necessary.)

hash_representations: dict

Dictionary of parameter names in the dataclass where you can provide functions that return a unique/consistent representation for each parameter to use as the value for hashing. Setting these allow control over what counts as a cache miss/cache hit. Sane defaults are applied to any parameters that don’t have a representation function specified here. (see hashing.get_parameter_hash_value)

Any function provided is expected to take self (the entire parameter set instance) and the value of the named parameter to be hashed.

You can also provide None for any parameters in order to exclude their value from the hash entirely. (This is useful for operational arguments e.g. the number of GPU’s or processes to run on, where you would not want changing that to invalidate the existing cached values.)

name: str = 'UNNAMED'

Parameter set name. This can be used to easily distinguish/refer to specific configurations in aggregate stages. This should be unique for every parameter set.

overwrite: bool = False

Whether to overwrite pre-cached values. Curifactory automatically sets this based on command line flags.

params_hash(dry=False)

Convenience function to see the hash of these parameters that curifactory is computing, or debug them with dry=True.