draccus 0.9.3

A slightly opinionated framework for simple dataclass-based configurations based on Pyrallis.

Draccus - Slightly Less Simple Configuration with Dataclasses

Draccus: "A large herbivorous reptilian creature, known for their ability to breathe fire."

Draccus is a fork of the excellent Pyrallis library, but with a few changes to make it more suitable for more complex use cases. The main changes are:

• Support for subtyping configs (that is, choosing between different configs based on a parameter)
• Support for including config files in config files
• Better support for containers of configs (e.g. a list of configs)

I swear I didn't want to fork it, but the Pyrallis devs (understandably) didn't want to merge some of these.

Why draccus?

We support everything in Pyrallis (see their examples), but also support subtyping and including config files within config files. We try to maintain the original repository's simple, clean approach.

With draccus your configuration is linked directly to your pre-defined dataclass, allowing you to easily create different configuration structures, including nested ones, using an object-oriented design. The parsed arguments are used to initialize your dataclass, inheriting the corresponding type hints and code completion features.

My First Draccus Example

(This example is the same as in Pyrallis. Draccus differs mainly in advanced features like subtyping.)

Here's a simple example of how to use draccus to parse arguments into a dataclass:

from dataclasses import dataclass
import draccus


@dataclass
class TrainConfig:
    """Training Config for Machine Learning"""
    workers: int = 8               # The number of workers for training
    exp_name: str = 'default_exp'  # The experiment name


@draccus.wrap()
def main(cfg: TrainConfig):
    print(f"Training {cfg.exp_name} with {cfg.workers} workers...")

The arguments can then be specified using command-line arguments, a yaml configuration file, or both.

$ python train_model.py --config_path=some_config.yaml --exp_name=my_first_exp
Training my_first_exp with 42 workers...

Assuming the following configuration file

exp_name: my_yaml_exp
workers: 42

model:
  type: bert
  num_layers: 24
  num_heads: 24
  hidden_size: 1024
  dropout: 0.2

Inclusion of Config Files

(This is a difference from Pyrallis.)

We support including config files from other config files via pyyaml-include. This is useful for splitting up your config into multiple files, or for including a base config file in your config.

It works like this:

# model_config.yaml
type: bert
num_layers: 24
num_heads: 24
hidden_size: 1024
dropout: 0.2

# train_config.yaml
exp_name: my_yaml_exp
workers: 42
model: !include model_config.yaml

Inclusion of config files using Command Line Arguments using include keyword

You can use include keyword in the command line arguments to include a config file in a nested way.

It works like this:

# model_config.yaml
type: bert_cli
num_layers: 48
num_heads: 48

# train_config.yaml
exp_name: my_yaml_exp
workers: 42
model: 
  type: bert
  num_layers: 24
  num_heads: 24

Using python train_model.py --config_path=train_config.yaml --model="include model_config.yaml" will give cfg.model.type = 'bert_cli'

Including Configs at Top Level

PyYAML, upon which draccus is based, supports a common YAML extension << for merging keys from multiple maps. We can combine this with !include to include a config file:

# base_config.yaml
type: bert
lr: 0.001

# train_config.yaml
<<: !include base_config.yaml
exp_name: my_yaml_exp

(I don't love this syntax, but it's consistent with PyYAML.)

More Flexible Configuration with Choice Types

(This is a difference from Pyrallis.)

Choice Types, aka "Sum Types" or "Tagged Unions", are a powerful way to define a choice of types that can be selected at runtime. For instance, you might want to choose what kind of model to train, or what kind of optimizer to use.

Draccus provides a ChoiceRegistry class that lets you define a choice of types that can be selected at runtime. You can then use the register_subclass decorator to register a subclass of your choice type. The type field of the choice type is used to select the subclass.

Here's a modified version of the example above, where we use a ChoiceRegistry to define a choice of model types:

from dataclasses import dataclass
import draccus


# Choice Registry lets you define a choice of implementations that can be selected at runtime
@dataclass
class ModelConfig(draccus.ChoiceRegistry):
    pass


@ModelConfig.register_subclass('gpt')
@dataclass
class GPTConfig(ModelConfig):
    """GPT Model Config"""
    num_layers: int = 12
    num_heads: int = 12
    hidden_size: int = 768


@ModelConfig.register_subclass('bert')
@dataclass
class BERTConfig(ModelConfig):
    """BERT Model Config"""
    num_layers: int = 12
    num_heads: int = 12
    hidden_size: int = 768
    dropout: float = 0.1


@dataclass
class TrainConfig:
    """Training Config for Machine Learning"""
    workers: int = 8                  # The number of workers for training
    exp_name: str = 'default_exp'     # The experiment name

    model: ModelConfig = GPTConfig()  # The model configuration


@draccus.wrap()
def main(cfg: TrainConfig):
    print(f"Training {cfg.exp_name} with {cfg.workers} workers...")

The arguments can then be specified using command-line arguments, a yaml configuration file, or both.

$ python train_model.py --config_path=some_config.yaml --exp_name=my_first_exp
Training my_first_exp with 42 workers...

Assuming the following configuration file

exp_name: my_yaml_exp
workers: 42

model:
  type: bert
  num_layers: 24
  num_heads: 24
  hidden_size: 1024
  dropout: 0.2

Everything below here is from Pyrallis. I'll update it eventually.

(It all still applies, substituting draccus for pyrallis.)

Key Features

Building on that design pyrallis offers some really enjoyable features including

• Builtin IDE support for autocompletion and linting thanks to the structured config. 🤓
• Joint reading from command-line and a config file, with support for specifying a default config file. 😍
• Support for builtin dataclass features, such as __post_init__ and @property 😁
• Support for nesting and inheritance of dataclasses, nested arguments are automatically created! 😲
• A magical @pyrallis.wrap() decorator for wrapping your main class 🪄
• Easy extension to new types using pyrallis.encode.register and pyrallis.decode.register 👽
• Easy loading and saving of existing configurations using pyrallis.dump and pyrallis.load 💾
• Magical --help creation from dataclasses, taking into account the comments as well! 😎
• Support for multiple configuration formats (yaml, json,toml) using pyrallis.set_config_type ⚙️

Getting to Know The pyrallis API in 5 Simple Steps 🐲

The best way to understand the full pyrallis API is through examples, let's get started!

🐲 1/5 pyrallis.parse for dataclass Parsing 🐲

Creation of an argparse configuration is really simple, just use pyrallis.parse on your predefined dataclass.

from dataclasses import dataclass, field
import draccus


@dataclass
class TrainConfig:
    """ Training config for Machine Learning """
    # The number of workers for training
    workers: int = field(default=8)
    # The experiment name
    exp_name: str = field(default='default_exp')


def main():
    cfg = draccus.parse(config_class=TrainConfig)
    print(f'Training {cfg.exp_name} with {cfg.workers} workers...')


if __name__ == '__main__':
    main()

Not familiar with dataclasses? you should probably check the Python Tutorial and come back here.

The config can then be parsed directly from command-line

$ python train_model.py --exp_name=my_first_model
Training my_first_model with 8 workers...

Oh, and pyrallis also generates an --help string automatically using the comments in your dataclass 🪄

$ python train_model.py --help
usage: train_model.py [-h] [--config_path str] [--workers int] [--exp_name str]

optional arguments:
  -h, --help      show this help message and exit
  --config_path str    Path for a config file to parse with pyrallis (default:
                  None)

TrainConfig:
   Training config for Machine Learning

  --workers int   The number of workers for training (default: 8)
  --exp_name str  The experiment name (default: default_exp)

🐲 2/5 The pyrallis.wrap Decorator 🐲

Don't like the pyrallis.parse syntax?

def main():
    cfg = pyrallis.parse(config_class=TrainConfig)
    print(f'Training {cfg.exp_name} with {cfg.workers} workers...')

One can equivalently use the pyrallis.wrap syntax 😎

@pyrallis.wrap()
def main(cfg: TrainConfig):
    # The decorator automagically uses the type hint to parsers arguments into TrainConfig
    print(f'Training {cfg.exp_name} with {cfg.workers} workers...')

We will use this syntax for the rest of our tutorial.

🐲 3/5 Better Configs Using Inherent dataclass Features 🐲

When using a dataclass we can add additional functionality using existing dataclass features, such as the post_init mechanism or @properties :grin:

from dataclasses import dataclass, field
from pathlib import Path
from typing import Optional
import draccus


@dataclass
class TrainConfig:
    """ Training config for Machine Learning """
    # The number of workers for training
    workers: int = field(default=8)
    # The number of workers for evaluation
    eval_workers: Optional[int] = field(default=None)
    # The experiment name
    exp_name: str = field(default='default_exp')
    # The experiment root folder path
    exp_root: Path = field(default=Path('/share/experiments'))

    def __post_init__(self):
        # A builtin method of dataclasses, used for post-processing our configuration.
        self.eval_workers = self.eval_workers or self.workers

    @property
    def exp_dir(self) -> Path:
        # Properties are great for arguments that can be derived from existing ones
        return self.exp_root / self.exp_name


@draccus.wrap()
def main(cfg: TrainConfig):
    print(f'Training {cfg.exp_name}...')
    print(f'\tUsing {cfg.workers} workers and {cfg.eval_workers} evaluation workers')
    print(f'\tSaving to {cfg.exp_dir}')

$ python -m train_model.py --exp_name=my_second_exp --workers=42
Training my_second_exp...
    Using 42 workers and 42 evaluation workers
    Saving to /share/experiments/my_second_exp

Notice that in all examples we use the explicit dataclass.field syntax. This isn't a requirement of pyrallis but rather a style choice. As some of your arguments will probably require dataclass.field (mutable types for example) we find it cleaner to always use the same notation.

🐲 4/5 Building Hierarchical Configurations 🐲

Sometimes configs get too complex for a flat hierarchy 😕, luckily pyrallis supports nested dataclasses 💥

@dataclass
class ComputeConfig:
    """ Config for training resources """
    # The number of workers for training
    workers: int = field(default=8)
    # The number of workers for evaluation
    eval_workers: Optional[int] = field(default=None)

    def __post_init__(self):
        # A builtin method of dataclasses, used for post-processing our configuration.
        self.eval_workers = self.eval_workers or self.workers


@dataclass
class LogConfig:
    """ Config for logging arguments """
    # The experiment name
    exp_name: str = field(default='default_exp')
    # The experiment root folder path
    exp_root: Path = field(default=Path('/share/experiments'))

    @property
    def exp_dir(self) -> Path:
        # Properties are great for arguments that can be derived from existing ones
        return self.exp_root / self.exp_name

# TrainConfig will be our main configuration class.
# Notice that default_factory is the standard way to initialize a class argument in dataclasses

@dataclass
class TrainConfig:
    log: LogConfig = field(default_factory=LogConfig)
    compute: ComputeConfig = field(default_factory=ComputeConfig)

@pyrallis.wrap()
def main(cfg: TrainConfig):
    print(f'Training {cfg.log.exp_name}...')
    print(f'\tUsing {cfg.compute.workers} workers and {cfg.compute.eval_workers} evaluation workers')
    print(f'\tSaving to {cfg.log.exp_dir}')

The argument parse will be updated accordingly

$ python train_model.py --log.exp_name=my_third_exp --compute.eval_workers=2
Training my_third_exp...
    Using 8 workers and 2 evaluation workers
    Saving to /share/experiments/my_third_exp

🐲 5/5 Easy Serialization with pyrallis.dump 🐲

As your config get longer you will probably want to start working with configuration files. Pyrallis supports encoding a dataclass configuration into a yaml file 💾

The command pyrallis.dump(cfg, open('run_config.yaml','w')) will result in the following yaml file

compute:
  eval_workers: 2
  workers: 8
log:
  exp_name: my_third_exp
  exp_root: /share/experiments

pyrallis.dump extends yaml.dump and uses the same syntax.

Configuration files can also be loaded back into a dataclass, and can even be used together with the command-line arguments.

cfg = pyrallis.parse(config_class=TrainConfig,
                              config_path='/share/configs/config.yaml')

# or the decorator synrax
@pyrallis.wrap(config_path='/share/configs/config.yaml')

# or with the CONFIG argument
python my_script.py --log.exp_name=readme_exp --config_path=/share/configs/config.yaml

# Or if you just want to load from a .yaml without cmd parsing
cfg = pyrallis.load(TrainConfig, '/share/configs/config.yaml')

Command-line arguments have a higher priority and will override the configuration file

Finally, one can easily extend the serialization to support new types 🔥

# For decoding from cmd/yaml
pyrallis.decode.register(np.ndarray,np.asarray)

# For encoding to yaml
pyrallis.encode.register(np.ndarray, lambda x: x.tolist())

# Or with the wrapper version instead
@pyrallis.encode.register
def encode_array(arr : np.ndarray) -> str:
    return arr.tolist()

🐲 That's it you are now a pyrallis expert! 🐲

Why Another Parsing Library?

XKCD 927 - Standards

The builtin argparse has many great features but is somewhat outdated :older_man: with one its greatest weakness being the lack of typing. This has led to the development of many great libraries tackling different weaknesses of argparse (shout out for all the great projects out there! You rock! :metal:).

In our case, we were looking for a library that would support the vanilla dataclass without requiring dedicated classes, and would have a loading interface from both command-line and files. The closest candidates were hydra and simple-parsing, but they weren't exactly what we were looking for. Below are the pros and cons from our perspective:

Hydra

A framework for elegantly configuring complex applications from Facebook Research.

• Supports complex configuration from multiple files and allows for overriding them from command-line.
• Does not support non-standard types, does not play nicely with datclass.__post_init__and requires a ConfigStore registration.

SimpleParsing

A framework for simple, elegant and typed Argument Parsing by Fabrice Normandin

• Strong integration with argparse, support for nested configurations together with standard arguments.
• No support for joint loading from command-line and files, dataclasses are still wrapped by a Namespace, requires dedicated classes for serialization.

We decided to create a simple hybrid of the two approaches, building from SimpleParsing with some hydra features in mind. The result, pyrallis, is a simple library that that is relatively low on features, but hopefully excels at what it does.

If pyrallis isn't what you're looking for we strongly advise you to give hydra and simpleParsing a try (where other interesting option include click, ext_argpase, jsonargparse, datargs and tap). If you do :heart: pyrallis then welcome aboard! We're gonna have a great journey together! 🐲

Tips and Design Choices

Beware of Mutable Types (or use pyrallis.field)

Dataclasses are great (really!) but using mutable fields can sometimes be confusing. For example, say we try to code the following dataclass

@dataclass
class OptimConfig:
    worker_inds: List[int] = []
    # Or the more explicit version
    worker_inds: List[int] = field(default=[])

As [] is mutable we would actually initialize every instance of this dataclass with the same list instance, and thus is not allowed. Instead dataclasses would direct you the default_factory function, which calls a factory function for generating the field in every new instance of your dataclass.

worker_inds: List[int] = field(default_factory=list)

Now, this works great for empty collections, but what would be the alternative for

worker_inds: List[int] = field(default=[1,2,3])

Well, you would have to create a dedicated factory function that regenerates the object, for example

worker_inds: List[int] = field(default_factory=lambda : [1,2,3])

Kind of annoying and could be confusing for a new guest reading your code :confused: Now, while this isn't really related to parsing/configuration we decided it could be nice to offer a sugar-syntax for such cases as part of pyrallis

from draccus import field

worker_inds: List[int] = field(default=[1, 2, 3], is_mutable=True)

The pyrallis.field behaves like the regular dataclasses.field with an additional is_mutable flag. When toggled, the default_factory is created automatically, offering the same functionally with a more reader-friendly syntax.

Uniform Parsing Syntax

For parsing files we opted for yaml as our format of choice, following hydra, due to its concise format. Now, let us assume we have the following .yaml file which yaml successfully handles:

compute:
  worker_inds: [0,2,3]

Intuitively we would also want users to be able to use the same syntax

python my_app.py --compute.worker_inds=[0,2,3]

However, the more standard syntax for an argparse application would be

python my_app.py --compute.worker_inds 0 2 3

We decided to use the same syntax as in the yaml files to avoid confusion when loading from multiple sources.

Not a yaml fun? pyrallis also supports json and toml formats using pyrallis.set_config_type('json') or with pyrallis.config_type('json'):

TODOs:

• Fix error with default Dict and List

    Underlying error: No decoding function for type ~KT, consider using pyrallis.decode.register

• Refine the --help command

For example the options argument is confusing there

• Add a test to omit_defaults

Contributors ✨

Thanks goes to these wonderful people (emoji key):

Ido Weiss 🎨 🤔

Yair Feldman 🎨 🤔

This project follows the all-contributors specification. Contributions of any kind welcome!