datenwissenschaften 1.0.0

A powerful Reinforcement Learning toolkit for classic games, featuring the Retro Speedlab Command Center.

datenwissenschaften: Retro Speedlab 🚀

datenwissenschaften is a high-performance Reinforcement Learning (RL) toolkit designed to streamline training on classic video games. Built on top of stable-baselines3 and stable-retro, it provides a robust framework for developing, monitoring, and perfecting RL agents.

At its core is the Retro Speedlab Command Center, a beautiful real-time dashboard that keeps you in control of your training sessions.

✨ Key Features

• 🎮 Retro Speedlab Command Center: A rich, interactive terminal dashboard to monitor steps, win rates, episodes per minute, and uptime in real-time.
• 🏋️ Effortless Training: Use the Trainer class to orchestrate complex training sessions with minimal boilerplate.
• 📊 Smart Callbacks: Automatic model checkpointing, best-episode recording (including .bk2 replays), and episode uploading.
• 🛠️ Robust Infrastructure: Built-in support for environment wrappers, ROM management, and flexible configuration via environment variables.
• 🚀 Optimized for Stable Retro: Seamless integration with the latest classic gaming environments.

🚀 Quick Start

Installation

pip install datenwissenschaften

Basic Usage

import os
from datenwissenschaften import Trainer, EnvironmentBuilder, ModelBuilder
from stable_baselines3 import PPO

# 1. Configure your environment
os.environ["RETRO_SPEEDLAB_GAME_ID"] = "SonicTheHedgehog-Genesis"
os.environ["RETRO_SPEEDLAB_ROM_PATH"] = "./roms"
os.environ["RETRO_SPEEDLAB_MODEL_DIR"] = "./models"
os.environ["RETRO_SPEEDLAB_RECORDING_DIR"] = "./recordings"
os.environ["RETRO_SPEEDLAB_TIMESTEPS"] = "1000000"

# 2. Build the environment (using a wrapper if needed)
def my_wrapper(env):
    return env

builder = EnvironmentBuilder(wrapper=my_wrapper)
venv = builder.build(n_envs=8)

# 3. Initialize model and trainer
model_builder = ModelBuilder(build_model=lambda env: PPO("CnnPolicy", env))
model = model_builder.build(venv)
trainer = Trainer()

# 4. Start training and watch the Command Center take off!
trainer.train(model)

⚙️ Configuration

The toolkit is highly configurable via environment variables:

Variable	Description	Default
RETRO_SPEEDLAB_GAME_ID	The ID of the game to train on (e.g., SonicTheHedgehog-Genesis).	Required
RETRO_SPEEDLAB_ROM_PATH	Path to the directory containing your ROMs.	Required
RETRO_SPEEDLAB_MODEL_DIR	Where to save trained models and checkpoints.	Required
RETRO_SPEEDLAB_RECORDING_DIR	Where to store episode recordings and .bk2 files.	Required
RETRO_SPEEDLAB_TIMESTEPS	Total number of timesteps to train for.	Required
RETRO_SPEEDLAB_NUM_ENVS	Number of parallel environments to run.	1
RETRO_SPEEDLAB_SAVESTATE	Specific savestate to start from.	None

📜 License

This project is licensed under the GNU General Public License v3.0. See the LICENSE file for details.

---

Developed with ❤️ by datenwissenschaften.