How yumizu-da/atmacup17 works
Main Components
• Docker & Development Environment:
• Dockerfile, compose.yaml, .devcontainer directory, and related configuration files.
• These components ensure that experiments run in a consistent and reproducible environment.
• Experiment Management:
• The “exp” directory holds multiple experiment subdirectories. Each experiment (e.g., “001”, “014”, “015”, etc.) contains:
◦ A “config” folder with YAML configuration files (including subdirectories for models that define specific models like “cb.yaml”, “lgb.yaml”, “bert.yaml”, “gemma.yaml”).
◦ Jupyter notebooks (e.g., “run.ipynb”, “eda.ipynb”) where experiments are executed.
◦ A “src” module housing supporting scripts (e.g., “seed.py”, “groupkfold.py”) to manage reproducibility and experiment-specific logic.
• That the naming and structure are used to manage different experiments and maintain reproducibility.
• Data & Results:
• The “data” directory (with placeholder .gitkeep) is for source data.
• The “results” directory is intended to store outputs from experiments.
• Dependency & Library Management:
• The presence of “pyproject.toml”, “.python-version”, and the use of the “uv” tool (as detailed in the README) indicate that dependency management is handled systematically.
• Library management flow from “uv add/remove” commands syncing with pyproject.toml.
Relationships and Interactions
• Docker and Environment:
• Experiment Execution Flow:
• Each experiment directory is completely self-contained: it holds its own configuration, experiment-specific code, and notebooks.
• Dependency Management:
• Role of “uv” syncing the libraries as a preparatory step before running experiments.
• Connection between “pyproject.toml” and other configuration components.
Architectural Patterns and Design Principles
• Reproducibility:
• The design emphasizes reproducibility; each experiment encapsulated in its own directory maintains environment-specific configurations (YAML files) and seed settings.
• Modular Experimentation:
• The splitting of experiments into independent directories under “exp” signifies a modular approach to experimentation.
• Containerization:
• Docker ensures consistency. The system is designed such that the environment (via the Docker/compose setup) is decoupled from the experiment specifics.
• Infrastructure as Code:
• The use of configuration files and environment descriptors (.devcontainer, compose.yaml) reflects best practices in managing environment dependencies.