Hooksett

A Python library providing a flexible, extensible hook system for managing parameters, metrics, and artifacts in machine learning workflows.

Hooksett is a Python library that provides a flexible, extensible hook system for managing parameters, metrics, and artifacts in ML workflows.

Example

Annotate your ML class with tracked types, wire up a config loader and MLflow output, and every parameter and metric is automatically captured:

from hooksett import tracked, HookManager
from hooksett.hooks import YAMLConfigInput, TypeValidationHook, MLflowOutput

type Parameter[T] = T
type Metric[T] = T

@tracked
class Trainer:
    learning_rate: Parameter[float] = 0.01
    batch_size: Parameter[int] = 32
    epochs: Parameter[int] = 100
    accuracy: Metric[float] = 0.0
    loss: Metric[float] = 0.0

    def train(self):
        for epoch in range(self.epochs):
            # training step ...
            self.accuracy = evaluate(model)
            self.loss = compute_loss(model)

# load params from YAML, validate, and log everything to MLflow
manager = HookManager()
manager.add_input_hook(YAMLConfigInput("config.yaml"))
manager.add_input_hook(TypeValidationHook())
manager.add_output_hook(MLflowOutput())

Features

• @tracked class decorator — monitors attribute changes on class instances via Python descriptors
• @track_function decorator — tracks function parameters and local variables; values are saved to hooks once at function/method exit
• Local variable tracking — annotate locals with Traced[T] inside methods or functions; only the final value at exit is captured
• Automatic parameter loading — YAMLConfigInput hook loads configuration from YAML files into tracked attributes
• Parameter validation — TypeValidationHook enforces type annotations; RangeValidationHook checks numeric bounds
• Custom type registry — define domain-specific tracked types (Parameter, Metric, Artifact, Prompt, Response, Feature) via register_tracked_type
• Pluggable output hooks — TracedOutput for logging, MLflowOutput for experiment tracking, or write your own OutputHook
• Singleton HookManager — register input and output hooks once; all decorated classes and functions use them automatically
• Separation of config and code — parameters live in YAML, validation in hooks, tracking in type annotations