Home

Awesome

Co-Rider

License Code style: black codecov

Tiny configuration library tailored for Deep Learning project and the Ride library.

pip install corider

Organising configurations and arguments for Deep Learning projects

Keeping track of, merging and exposing configurations as arguments can be cumbersome and introduces a lot of boiler-plate code. This tiny library aims to introduce a configuration structure, that will fit many Deep Learning projects.

A basic configuration is defined as follows:

from corider import Configs

c1 = Configs()
c1.add(
    name="learning_rate",
    type=int,
    default=2,
    strategy="loguniform",
    description="Learning rate for optimizer",
    choices=(1e-8, 1),
)
c1.add(
    name="optimizer",
    type=str,
    default="sgd",
    strategy="constant",
    description="Optimizer to use.",
    choices=["sgd", "adam"],
)

Argparse

Co-Rider is fully compartible with argparse and can both load and dump argparse configurations:

# argparse_example.py
from argparse import ArgumentParser
from corider import Configs

parser = ArgumentParser(add_help=True)
parser.add_argument(
    "--defined_with_argparse",
    default=42,
    choices=(42, 1337),
    type=int,
    help="Nonsensical parameter defined for demo purposes.",
)

c2 = Configs.from_argument_parser(parser)

c2.add(
    name="defined_with_corider",
    type=int,
    default="lit",
    description="Another parameter for demo purposes",
    choices=["lit", "woke"],
)

new_parser = c2.add_argparse_args(ArgumentParser(add_help=True))

args = new_parser.parse_args()

# Do somethin with the args

Use from shell as usual:

$ python argparse_example.py --help
usage: argparse_example.py [-h] [--defined_with_argparse {42,1337}]
                           [--defined_with_corider {lit,woke}]

optional arguments:
  -h, --help            show this help message and exit
  --defined_with_argparse {42,1337}
                        Nonsensical parameter defined for demo purpose.
                        (Default: 42)
  --defined_with_corider {lit,woke}
                        Another parameter for demo purpose (Default: lit)

Ray Tune

By now you may have wodered about the strategy parameter. This parameter is intended for hyperparameter optimizers to indicate which sampling strategy to employ during hyperparameter search.

Four strategies are available:

For now, an automatic export to Ray[Tune] is included, which can be used as follows:

from ray import tune

# Configs which had strategy "constant" can be added as argparse args
parser = c.add_tune_argparse_args(ArgumentParser())
args = parser.parse_args()

# Other parameters are exported in a Tune-compatible format
tune_config = c.tune_config()

# Run search
analysis = tune.run(
    your_training_function,
    config=tune_config,
    ... # Other tune.run parameters
)

Argument addition and subtracktion

Co-Rider can add and subtract configs as needed:

c1 = ...  # As defined above (has: "learning_rate", "optimizer")

c2 = ...  # As defined above (has: "defined_with_argparse", "defined_with_corider")

c3 = Configs()
c3.add(
    name="learning_rate",  # Also defined in c1
    type=int,
    default=2,
    strategy="loguniform",
    description="Learning rate for optimizer",
    choices=(1e-8, 1),
)

# Has: "optimizer," "defined_with_argparse", "defined_with_corider"
c4 = c1 + c2 - c3  

Load configuration from file

A configuration can be loaded from either .yaml or .json formatted files:

# example_conf.yaml
dropout:
  type: float
  strategy: choice
  choices: [0.0, 0.1, 0.2, 0.3, 0.4]
learning_rate:
  type: float
  strategy: loguniform
  choices: [0.01, 0.5]
weight_decay:
  type: float
  strategy: loguniform
  choices: [0.000001, 0.001]
c = Configs.from_file("example_conf.yaml")