Configurize

A Python library for hierarchical configuration management with inheritance, cross-references, and diffing support.

Installation

pip install configurize

for development

pip install .

Features

• Hierarchical configs: Nest configs within configs with automatic parent-child relationships
• Inheritance: Extend configs like regular Python classes
• Cross-references: Use Ref() to reference values from other parts of the config tree
• Diffing: Compare two configs and see what changed
• Merge: Apply updates from dicts or other configs
• Validation: Define critical_keys for consistency checks

Quick Start

from configurize import Config, Ref

class ModelConfig(Config):
    in_channels = 32
    out_channels = 64

class TrainerConfig(Config):
    train_iters = 100
    batch_size = 16

class Exp(Config):
    model = ModelConfig
    trainer = TrainerConfig

# Create and use
exp = Exp()
print(exp.model.in_channels)  # 32
print(exp.trainer.train_iters)  # 100

Cross-References with Ref

Use Ref() to reference values from other parts of the config tree:

class ModelConfig(Config):
    hidden_dim = Ref('..hidden_dim')  # Reference parent's hidden_dim

class Exp(Config):
    hidden_dim = 256
    model = ModelConfig

exp = Exp()
print(exp.model.hidden_dim)  # 256 (resolved from parent)

Reference syntax:

• .attr - self.attr
• ..attr - parent.attr
• ...sub.attr - grandparent.sub.attr

Merging and Overrides

exp = Exp()

# Merge from dict (supports dot notation)
exp.merge({'model.in_channels': 64, 'trainer.train_iters': 200})

# Temporary modifications
with exp.trainer.modify(batch_size=32):
    print(exp.trainer.batch_size)  # 32
print(exp.trainer.batch_size)  # 16 (restored)

Diffing Configs

exp1 = Exp()
exp2 = Exp()
exp2.model.in_channels = 128

diff = exp1.diff(exp2)
print(diff)  # Shows differences with colors

CLI Tool

After installation, use cfshow to inspect and compare config files:

# Show a config
cfshow examples/train_example.py

# Compare two configs
cfshow base_exp.py new_exp.py

# Inspect a sub-config
cfshow my_exp.py --key=model

Example

See examples/train_example.py for a complete example showing:

• Nested config structure (logger, model, trainer)
• Builder pattern with build_* methods
• Running an experiment from config