⚙️ YAML configuration meets Python 🐍
Define Python objects in YAML. Reference, compose, and instantiate them effortlessly.
Stop hardcoding parameters. Define complex Python objects in clean YAML files, compose them naturally, and instantiate with one line.
# config.yaml
model:
_target_: torch.nn.Linear
in_features: 784
out_features: "%dataset::num_classes" # Reference other values
dataset:
num_classes: 10from sparkwheel import Config
config = Config.load("config.yaml")
model = config.resolve("model") # Actual torch.nn.Linear(784, 10) instance!- Declarative Object Creation - Instantiate any Python class from YAML with
_target_ - Smart References -
@for resolved values,%for raw YAML - Composition by Default - Configs merge naturally (dicts merge, lists extend)
- Explicit Operators -
=to replace,~to delete when needed - Python Expressions - Compute values dynamically with
$prefix - Schema Validation - Type-check configs with Python dataclasses
- CLI Overrides - Override any value from command line
pip install sparkwheelSparkwheel builds on similar ideas but adds powerful features:
| Feature | Hydra/OmegaConf | Sparkwheel |
|---|---|---|
| Config composition | Explicit (+, ++) |
By default (dicts merge, lists extend) |
| Replace semantics | Default | Explicit with = operator |
| Delete keys | Not idempotent | Idempotent ~ operator |
| References | OmegaConf interpolation | @ (resolved) + % (raw YAML) |
| Python expressions | Limited | Full Python with $ |
| Schema validation | Structured Configs | Python dataclasses |
| List extension | Lists replace | Lists extend by default |
Composition by default means configs merge naturally without operators:
# base.yaml
model:
hidden_size: 256
dropout: 0.1
# experiment.yaml
model:
hidden_size: 512 # Override
# dropout inherited- Discord Server - Chat with the community
- YouTube Channel - Tutorials and demos
- GitHub Issues - Bug reports and feature requests
We welcome contributions! See CONTRIBUTING.md for development setup and guidelines.
Sparkwheel is a hard fork of MONAI Bundle's configuration system, refined and expanded for general-purpose use. We're deeply grateful to the MONAI team for their excellent foundation.
Sparkwheel powers Lighter, our configuration-driven deep learning framework built on PyTorch Lightning.
Apache License 2.0 - See LICENSE for details.