collie-mlops 0.1.1b0

A Lightweight MLOps Framework for Machine Learning Workflows

Collie

A Lightweight MLOps Framework for Machine Learning Workflows

Overview

Collie is a modern MLOps framework designed to streamline machine learning workflows by providing a component-based architecture integrated with MLflow. It enables data scientists and ML engineers to build, deploy, and manage ML pipelines with ease through modular components that handle different stages of the ML lifecycle.

Features

• Component-Based Architecture: Modular design with specialized components for each ML workflow stage
• MLflow Integration: Built-in experiment tracking, model registration, and deployment capabilities
• Pipeline Orchestration: Seamless workflow management with event-driven architecture
• Model Management: Automated model versioning, staging, and promotion
• Framework Agnostic: Supports multiple ML frameworks (PyTorch, scikit-learn, XGBoost, LightGBM, Transformers)

Architecture

Collie follows an event-driven architecture with the following core components:

• Transformer: Data preprocessing and feature engineering
• Tuner: Hyperparameter optimization
• Trainer: Model training and validation
• Evaluator: Model evaluation and comparison
• Pusher: Model deployment and registration
• Orchestrator: Workflow coordination and execution

Quick Start

Installation

pip install collie-mlops

This will install Collie with all supported ML frameworks including:

• scikit-learn
• PyTorch
• XGBoost
• LightGBM
• Transformers (with Sentence Transformers)

Prerequisites

• Python >= 3.10
• MLflow tracking server (can be local or remote)

Components

Transformer

Handles data preprocessing, feature engineering, and data validation.

class CustomTransformer(Transformer):
    def handle(self, event) -> Event:
        # Process your data
        processed_data = ... 
        return Event(payload=TransformerPayload(train_data=processed_data))

Tuner

Performs hyperparameter optimization using various strategies.

class CustomTuner(Tuner):
    def handle(self, event) -> Event:
        # Optimize hyperparameters
        best_params = ...
        return Event(payload=TunerPayload(hyperparameters=best_params))

Trainer

Trains machine learning models with automatic experiment tracking.

class CustomTrainer(Trainer):
    def handle(self, event) -> Event:
        # Train your model
        model = ...
        return Event(payload=TrainerPayload(model=model))

Evaluator

Evaluates model performance and decides on deployment.

class CustomEvaluator(Evaluator):
    def handle(self, event) -> Event:
        # Evaluate model performance
        metrics = ...
        is_better: bool = ...
        return Event(payload=EvaluatorPayload(
            metrics=metrics, 
            is_better_than_production=is_better
        ))

Pusher

Handles model deployment and registration.

class CustomPusher(Pusher):
    def handle(self, event) -> Event:
        # Deploy model to production
        model_uri = ...
        return Event(payload=PusherPayload(model_uri=model_uri))

Configuration

MLflow Setup

Start MLflow tracking server:

mlflow server \
    --backend-store-uri sqlite:///mlflow.db \
    --default-artifact-root ./mlruns \
    --host 0.0.0.0 \
    --port 5000

Supported Frameworks

Collie supports multiple ML frameworks through its model flavor system currently:

• PyTorch
• scikit-learn
• XGBoost
• LightGBM
• Transformers

Documentation

Here you are

Roadmap

• TensorFlow/Keras support
• Model monitoring and drift detection
• Integrate an LLM training/fine-tuning framework

License

This project is licensed under the MIT License - see the LICENSE file for details.

Citation

If you use Collie in your research, please cite:

@software{collie2025,
  author = {ChingHuanChiu},
  title = {Collie: A Lightweight MLOps Framework},
  year = {2025},
  url = {https://github.com/ChingHuanChiu/collie}
}

---