README.md

Confection: The sweetest config system for Python

confection 🍬 is a lightweight library that offers a configuration system letting you conveniently describe arbitrary trees of objects.

Configuration is a huge challenge for machine-learning code because you may want to expose almost any detail of any function as a hyperparameter. The setting you want to expose might be arbitrarily far down in your call stack, so it might need to pass all the way through the CLI or REST API, through any number of intermediate functions, affecting the interface of everything along the way. And then once those settings are added, they become hard to remove later. Default values also become hard to change without breaking backwards compatibility.

To solve this problem, confection offers a config system that lets you easily describe arbitrary trees of objects. The objects can be created via function calls you register using a simple decorator syntax. You can even version the functions you create, allowing you to make improvements without breaking backwards compatibility. The most similar config system we’re aware of is Gin, which uses a similar syntax, and also allows you to link the configuration system to functions in your code using a decorator. confection's config system is simpler and emphasizes a different workflow via a subset of Gin’s functionality.

⏳ Installation

pip install confection

conda install -c conda-forge confection

👩‍💻 Usage

The configuration system parses a .cfg file like

[training]
patience = 10
dropout = 0.2
use_vectors = false

[training.logging]
level = "INFO"

[nlp]
# This uses the value of training.use_vectors
use_vectors = ${training.use_vectors}
lang = "en"

and resolves it to a Dict:

{
  "training": {
    "patience": 10,
    "dropout": 0.2,
    "use_vectors": false,
    "logging": {
      "level": "INFO"
    }
  },
  "nlp": {
    "use_vectors": false,
    "lang": "en"
  }
}

The config is divided into sections, with the section name in square brackets – for example, [training]. Within the sections, config values can be assigned to keys using =. Values can also be referenced from other sections using the dot notation and placeholders indicated by the dollar sign and curly braces. For example, ${training.use_vectors} will receive the value of use_vectors in the training block. This is useful for settings that are shared across components.

The config format has three main differences from Python’s built-in configparser:

1. JSON-formatted values. confection passes all values through json.loads to interpret them. You can use atomic values like strings, floats, integers or booleans, or you can use complex objects such as lists or maps.
2. Structured sections. confection uses a dot notation to build nested sections. If you have a section named [section.subsection], confection will parse that into a nested structure, placing subsection within section.
3. References to registry functions. If a key starts with @, confection will interpret its value as the name of a function registry, load the function registered for that name and pass in the rest of the block as arguments. If type hints are available on the function, the argument values (and return value of the function) will be validated against them. This lets you express complex configurations, like a training pipeline where batch_size is populated by a function that yields floats.

There’s no pre-defined scheme you have to follow; how you set up the top-level sections is up to you. At the end of it, you’ll receive a dictionary with the values that you can use in your script – whether it’s complete initialized functions, or just basic settings.

For instance, let’s say you want to define a new optimizer. You'd define its arguments in config.cfg like so:

[optimizer]
@optimizers = "my_cool_optimizer.v1"
learn_rate = 0.001
gamma = 1e-8

To load and parse this configuration using a catalogue registry (install catalogue separately):

import dataclasses
from typing import Union, Iterable
import catalogue
from confection import registry, Config

# Create a new registry.
registry.optimizers = catalogue.create("confection", "optimizers", entry_points=False)


# Define a dummy optimizer class.
@dataclasses.dataclass
class MyCoolOptimizer:
    learn_rate: float
    gamma: float


@registry.optimizers.register("my_cool_optimizer.v1")
def make_my_optimizer(learn_rate: Union[float, Iterable[float]], gamma: float):
    return MyCoolOptimizer(learn_rate, gamma)


# Load the config file from disk, resolve it and fetch the instantiated optimizer object.
config = Config().from_disk("./config.cfg")
resolved = registry.resolve(config)
optimizer = resolved["optimizer"]  # MyCoolOptimizer(learn_rate=0.001, gamma=1e-08)

⚠️ Caution: Type-checkers such as mypy will mark adding new attributes to registry this way - i. e. registry.new_attr = ... - as errors. This is because a new attribute is added to the class after initialization. If you are using typecheckers, you can either ignore this (e. g. with # type: ignore for mypy) or use a typesafe alternative: instead of registry.new_attr = ..., use setattr(registry, "new_attr", ...).

Under the hood, confection will look up the "my_cool_optimizer.v1" function in the "optimizers" registry and then call it with the arguments learn_rate and gamma. If the function has type annotations, it will also validate the input. For instance, if learn_rate is annotated as a float and the config defines a string, confection will raise an error.

The Thinc documentation offers further information on the configuration system:

• recursive blocks
• defining variable positional arguments
• using interpolation
• using custom registries
• advanced type annotations with Pydantic
• using base schemas
• filling a configuration with defaults

🎛️ API

class Config

This class holds the model and training configuration and can load and save the INI-style configuration format from/to a string, file or bytes. The Config class is a subclass of dict and uses Python’s ConfigParser under the hood.

^{method Config.__init__}