How google/dopamine works
Project type and purpose
• This is an open-source research framework for rapid prototyping of reinforcement-learning (RL) agents, called “Dopamine.”
• It provides battle-tested implementations of algorithms (DQN, Rainbow, IQN, SAC, PPO) in both TensorFlow (legacy) and JAX, plus experiment runners, replay buffers, metrics, plotting utilities, Colab notebooks, baselines data, Docker containers, and extensive documentation.
File structure
Top-level directories and purposes:
• dopamine/ → core Python package
– continuous_domains/ discrete_domains/ → train & run scripts for Gym/ALE/MuJoCo
– jax/ → JAX-based agents, networks, replay memory, checkpointers
– tf/ → TensorFlow-based agents, replay memory
– labs/ → experimental extensions (Atari-100k, MOEs, offline-RL, etc.)
– colab/ → Jupyter notebooks & helpers for interactive use
– metrics/ → collectors, dispatchers for console, TensorBoard, pickle
– utils/ → plotting, visualization, test utilities
• baselines/ → JSON baseline results & static plots for Atari & MuJoCo
• docs/ → Sphinx/Markdown source for API docs & user guides
• docker/ → Dockerfiles for core, Atari, and MuJoCo environments
• tests/ → unit/integration tests covering all subpackages
README for insights
• Confirms design goals (easy experimentation, compactness, reproducibility).
• Notes both TF (legacy) and JAX (actively maintained) implementations.
• Highlights use of Gin configuration files (.gin) to drive experiments.
• Provides pointers to Colab notebooks, baselines data, and Docker usage.