Ryn SDK

A unified Python SDK for data, model, and trainer workflows.
This repository mirrors the automatic release workflow used in model-registry (auto-publish to PyPI via GitHub Actions).

Default install pulls everything. Subsets are available via extras (ryn[model], ryn[data], ryn[trainer]).

Table of Contents

• Installation
• Quick Start
• Project Structure
• Requirements Layout
• Versioning & Auto-Publish to PyPI
• Local Development
• Update Checker (Optional)
• Supported Python Versions
• Contributing
• License

Installation

Default (everything): installs the union of data + model + trainer requirements.

# Everything
pip install ryn

# Only the model stack
pip install "ryn[model]"

# Only the data stack
pip install "ryn[data]"

# Only the trainer stack
pip install "ryn[trainer]"

Notes

• Some dependencies (e.g., torch) may require platform-specific wheels. If you're on a GPU machine, consult the official PyTorch install guide for your CUDA/ROCm toolkit.
• If you need a slim install instead, use a virtual environment and install only the extras you need.

Quick Start

import ryn
from ryn import data, model, trainer

print("Ryn version:", ryn.__version__)

# Example scaffolding (replace with your actual API calls):
# Dataset operations
# ds = data.load_dataset("your-dataset-id")
# df = data.to_pandas(ds)

# Model registry / storage operations
# model_id = model.register_model(name="my-model", version="1.0.0", metadata={"task": "classification"})
# meta = model.get_model(name="my-model", version="1.0.0")

# Training utilities
# run = trainer.train(config={"epochs": 10, "lr": 1e-3})
# metrics = trainer.evaluate(run)

The snippet above shows typical entry points. Replace with your actual functions/classes provided by ryn.data, ryn.model, and ryn.trainer in this repo.

Project Structure

ryn/
  data/       # Data utilities & loaders
  model/      # Model registry / storage / inference helpers
  trainer/    # Training loops, trackers, and helpers
  __init__.py # Exposes public API and package metadata

Requirements Layout

Dependencies are split into three groups under requirements/ and combined for the default install:

requirements/
  model.txt    # e.g., boto3, huggingface-hub, transformers, einops, ...
  data.txt     # e.g., numpy, pandas, scikit-learn, datasets, openml, ...
  trainer.txt  # e.g., torch, torchvision, timm, albumentations, ...

• Default install (pip install ryn) = union(model + data + trainer).
• Extras for explicit subsets:
  • ryn[model]
  • ryn[data]
  • ryn[trainer]

These groups are wired in setup.py so that PyPI metadata carries the dependency info.

Versioning & Auto-Publish to PyPI

The repository uses a CI workflow (GitHub Actions) that auto-bumps and publishes to PyPI on push to main when the following conditions are met:

1. Repository variable PUBLISH_TO_PYPI is set to true.
2. The commit message contains the phrase pipy commit -push.

Bump level is controlled via the commit message:

• pipy commit -push major → bump MAJOR
• pipy commit -push minor → bump MINOR
• pipy commit -push (no keyword) → bump PATCH (default)

Workflow outline

1. Read current version from PyPI.
2. Compute next version (patch/minor/major).
3. Update ryn/__init__.py and setup.py.
4. Build distributions (sdist + wheel).
5. Upload to PyPI using Twine.

Required repo settings

• Settings → Secrets and variables → Actions → Secrets
  • PYPI_API_TOKEN (PyPI token)
• Settings → Secrets and variables → Actions → Variables
  • PUBLISH_TO_PYPI = true

Local Development

# Clone
git clone https://github.com/AIP-MLOPS/rayen.git
cd rayen

# (Optional) Create a virtual environment
python -m venv .venv && source .venv/bin/activate  # or .venv\Scripts\activate on Windows

# Install in editable mode with a subset of extras, e.g., model
pip install -e ".[model]"

# Build locally
python -m pip install --upgrade pip build twine
python -m build
twine check dist/*

If you prefer a full dev install, use pip install -e ".[all]" (identical to default deps).

Update Checker (Optional)

The package may include a lightweight update checker that prints a hint if a newer version is available on PyPI. This check is best-effort and should never fail your import due to network issues.

Supported Python Versions

• Python 3.11+

Contributing

1. Fork the repository and create a feature branch.
2. Keep commits focused and meaningful.
3. Open a Pull Request with a clear description and, if applicable, tests or examples.

Consider running your linter/formatter locally (e.g., ruff) before opening a PR.

License

MIT