emkonfig

Emkonfig

This is a Hydra alternative for configuration management. It adds features that I found missing in Hydra, and removes some that I found unnecessary.

Installation

pip install emkonfig

Usage

Parse a configuration file

from emkonfig import Emkonfig

emkonfig = Emkonfig("examples/configs/config.yaml")
config = emkonfig.parse()
emkonfig.print(config)

As it is in Hydra, the config is a DictConfig object coming from OmegaConf.

Parsers

Reference YAML parser

You can reference a YAML file path in your configuration, and Emkonfig automatically parses it.

# config.yaml
head: ${./path/to/head.yaml}

# ./path/to/head.yaml
in_features: 512
out_features: 1024

Will be parsed as:

head:
  in_features: 512
  out_features: 1024

Reference key parser

You can reference a key in your configuration, and Emkonfig automatically parses it.

some_parameter: 3.14
some_dict_parameter:
  key1: ${some_parameter}
  key2: some_other_parameter

some_list_parameter:
  - item1
  - ${some_dict_parameter.key2}
  - item3

final_parameter: ${some_list_parameter[2]}

Will be parsed as:

some_parameter: 3.14
some_dict_parameter:
  key1: 3.14
  key2: some_other_parameter

some_list_parameter:
  - item1
  - some_other_parameter
  - item3

final_parameter: item3

Class slug parser

In Emkonfig, you can register your classes, then reference them using the slug you registered them with. You can either use emkonfig.registry.register or emkonfig.registry.register_class for this.

NOTE: You have to make sure that the module containing the classes you want to register is imported before parsing the configuration file. For that, you can use the utility function: emkonfig.utils.import_modules(dir_name: str, exclude: list[str] | set[str] | None = None, verbose: bool = False). This function imports all of the modules under the given directory. If you encounter a parsing error stating that the reference slug for a class couldn't be found, it it very likely that there is an error in one of your files.

# my_project.some_module.some_file.py
from emkonfig.registry import register, register_class

@register("my_class")
class MyClass:
    def __init__(self, a, b=3):
        self.a = a
        self.b = b


class MyOtherClass:
    def __init__(self, c, d):
        self.c = c
        self.d = d

register_class("my_other_class", MyOtherClass, c=2, d=4)
register_class("my_other_class_with_other_params", MyOtherClass, c=3, d=5)

Now you can reference these classes in your configuration:

_{my_class}:
  a: 1

_{my_class as some_other_name}:
  a: 2
  b: 4

some_key:
  _{my_class as _}:
    a: 3

some_other_key:
  - _{my_other_class}: null

_{my_other_class}: null # to use the default parameters
_{my_other_class_with_other_params}: null

Will be parsed as:

my_class:
  _target_: my_project.some_module.some_file.MyClass
  a: 1
  b: 3

some_other_name:
  _target_: my_project.some_module.some_file.MyClass
  a: 2
  b: 4

some_key:
  _target_: my_project.some_module.some_file.MyClass
  a: 3
  b: 3

some_other_key:
  - _target_: my_project.some_module.some_file.MyOtherClass
    c: 2
    d: 4

my_other_class:
  _target_: my_project.some_module.some_file.MyOtherClass
  c: 2
  d: 4

my_other_class_with_other_params:
  _target_: my_project.some_module.some_file.MyOtherClass
  c: 3
  d: 5

You can directly use these parameters to instantiate these classes:

from emkonfig.utils import instantiate
my_class = instantiate(config.my_class)

Arguments parser

As Hydra, Emkonfig supports arguments parsing. You can pass arguments to your configuration file using --overwrites:

# config.yaml
some_parameter: 3
some_other_parameter:
  key1: key1
  key2: 3.14

python ./my_project/entrypoint.py --overwrites some_paramter=4 some_other_parameter.key2=2.71

The final configuration will be:

some_parameter: 4
some_other_parameter:
  key1: key1
  key2: 2.71

Defaults list parser

As it is in Hydra, you can define a list of default values for a key in your configuration. The key refer to the relative directories to your main config file, and the values refer to the .yaml files in these directories. Let's say this is the structure of your configs directory

configs/
  - config.yaml
  - head/
    - linear.yaml
  - backbone/
    - vision_backbone/
      - resnet50.yaml
    - language_backbone/
      - bert.yaml
  - optimizer/
    - adam.yaml
  - callbacks/
    - early_stopping.yaml
    - model_checkpoint.yaml
    - tensorboard.yaml

You can reference these yaml files in your defaults list as follows:

# config.yaml

defaults:
  - head: linear
  - backbone/vision_backbone: resnet50
  - backbone/language_backbone: bert
  - backbone/vision_backbone@model: resnet50 # You can also rename the key using '@'
  - optimizer: adam
  - callbacks:
      - early_stopping
      - model_checkpoint
      - tensorboard

Will be parsed as:

head: # parsed from configs/head/linear.yaml
backbone:
  vision_backbone: # parsed from configs/backbone/vision_backbone/resnet50.yaml
  language_backbone: # parsed from configs/backbone/language_backbone/bert.yaml
model: # parsed from configs/backbone/vision_backbone/resnet50.yaml
optimizer: # parsed from configs/optimizer/adam.yaml
callbacks:
  -  # parsed from configs/callbacks/early_stopping.yaml
  -  # parsed from configs/callbacks/model_checkpoint.yaml
  -  # parsed from configs/callbacks/tensorboard.yaml

Examples

If you want to see an example configuration file, you can check the examples/configs/config.yaml file, and run:

python ./examples/example.py