ezconfy 0.2.0

YAML-based configuration with Pydantic validation and dynamic object instantiation.

Why EzConfy?

ML projects constantly deal with configuration: learning rates, model parameters, dataset options, augmentation pipelines. EzConfy gives you typed, validated configs with automatic object wiring — without the complexity of a full framework like Hydra.

from ezconfy import ConfigBuilder

cfg = ConfigBuilder.from_files(config_paths="config.yaml", schema_path="schema.yaml")
print(cfg.training.batch_size)  # validated, typed access

What you get

Installation

pip install ezconfy

Requires Python 3.11+.

Quick Start

config.yaml

lr: 0.001
batch_size: 32
data_path: ./data

schema.yaml

lr: float
batch_size: int
data_path: pathlib:Path

train.py

from ezconfy import ConfigBuilder

cfg = ConfigBuilder.from_files(
    config_paths="config.yaml",
    schema_path="schema.yaml",
)

print(cfg.lr)          # 0.001 (float)
print(cfg.batch_size)  # 32 (int)
print(cfg.data_path)   # PosixPath('data') — automatically cast

Schema

A schema file describes the expected shape and types of your configuration:

types:
  OptimizerType:
    - adam
    - sgd
    - rmsprop

schema:
  model:
    hidden_dims: list[int]
  training:
    batch_size: int = 32
    num_epochs: int = 10
    shuffle: bool = true
    dropout: float?
    optimizer: OptimizerType

If no types are needed, the entire YAML is treated as the root schema (no schema: wrapper required).

Object Instantiation

Construct Python objects directly from config using _target_type_:

dataset:
  _target_type_: mypackage.data:MyDataset
  _init_args_:
    num_classes: 100
    root: /data

Use _init_method_ for alternative constructors (e.g. from_pretrained):

encoder:
  _target_type_: mypackage.models:BertEncoder
  _init_method_: from_pretrained
  _init_args_:
    model_name: bert-base-uncased

Placeholders & Expressions

Reference other config values with ${key}. Supports attribute access, method calls, and arithmetic:

lr: 0.001
warmup_lr: ${lr * 10}                       # arithmetic

num_classes: 10

dataset:
  _target_type_: mypackage.data:MyDataset
  _init_args_:
    num_classes: ${num_classes}               # scalar reference

model:
  _target_type_: mypackage.models:Classifier
  _init_args_:
    in_features: ${dataset.num_classes}       # attribute access
    params: ${encoder.parameters()}           # method call

Objects are instantiated in topological order based on their dependencies — forward references work automatically.

Multi-file Configs & Overrides

Pass multiple files — they are deep-merged in order (later files win on conflicts):

cfg = ConfigBuilder.from_files(
    config_paths=["base.yaml", "experiment.yaml"],
)

Apply programmatic overrides on top:

cfg = ConfigBuilder.from_files(
    config_paths="config.yaml",
    overrides={"training": {"batch_size": 64}},
)

Patch a single list element instead of restating the whole list — ... keeps the other elements:

callbacks:
  - ...
  - _id_: early_stopping
    _init_args_:
      patience: 20

Code Generation CLI

Generate a Pydantic model file from a schema for editor autocompletion and static analysis:

ezconfy generate schema.yaml -o models.py

Documentation

Full documentation: alessioarcara.github.io/EzConfy

License

MIT