How wywywang/coachai-projects works
Project Type and Purpose
• This is a single “monorepo” that bundles nine distinct badminton‐analytics sub-projects (e.g., Visualization Platform, Shot Influence, Stroke Forecasting, Strategic Environment, Movement Forecasting, CoachAI Environment, RallyNet, ShuttleSet, CoachAI-Challenge).
• Each sub-project is essentially an ML/data-processing pipeline or an RL environment plus a front-end (Visualization Platform). They run end-to-end experiments: ingest raw match data → preprocess → model training/prediction → evaluation → (optionally) visualization.
High-Level Components
A. Data Sources (raw CSVs, datasets)
• ShuttleSet, ShuttleSet22, RallyNet/Data, Strategic Environment/Data, Stroke Forecasting/data, CoachAI-Challenge data, etc.
B. Preprocessing Layer
• Python scripts/notebooks: badmintoncleaner.py, data_cleaner.py, prepare_dataset.py, preprocess scripts under each sub-project, Visual Platform’s preprocessing code.
C. Modeling / Training Pipelines
1. Shot Influence Module (Shot Influence/)
– cnn.py, rally_classifier.py, train.py, evaluate.py, util.py
2. Stroke Forecasting Module (Stroke Forecasting/)
– ShuttleNet, LSTM, Transformer subfolders + train.py, evaluate.py, generator.py, utils.py
3. Movement Forecasting Module (Movement Forecasting/)
– DyMF, rGCN, GCN, LSTM, Transformer code + train.py, evaluate.py, runner scripts
4. Strategic Environment & CoachAI RL Env (Strategic Environment/, CoachAI Badminton Environment/)
– multiagent core, environment, policy, Gym-style wrappers, RL algorithms (a2c.py, ppo.py, ddpg.py)
5. RallyNet Module (RallyNet/)
– data loaders, policy/models (generator, transformer), train.py, evaluate.py
D. Evaluation & Results
– evaluate.py in each module, saved_model directories (RallyNet/Results), demo.csv, metrics scripts
E. Visualization Front-End
– Visualization Platform/cgi-bin (CGI-Python scripts), HTML/CSS/JS pages, autoanalysis.html, ourdata.html, trajectory.html
Relationships & Data Flows
• Preprocessing → Modeling: cleaned datasets feed into each module’s train.py/runner.py.
• Evaluation → Visualization (optional): some evaluation outputs (e.g. prediction CSVs, heatmaps) are ingested by the Visualization Platform.
Architectural Patterns & Principles
• Modular pipeline pattern: clear separation between data, preprocessing, modeling, evaluation, and visualization.
• Reusable components: each sub-project implements its own environment.yml and isolated codebase but follows the same flow.
• Layered architecture: Data layer → Processing layer → Application (model) layer → Presentation (visualization) layer.
Technologies & Frameworks
• Language: Python 3.x (+ Jupyter notebooks)
• ML Frameworks: PyTorch/TensorFlow (implicit in model files), scikit-learn (for classical models in RallyNet), custom CNN/Transformer code
• RL: OpenAI Gym–style multiagent environment, stable-baseline-style algorithms (ppo.py, a2c.py, ddpg.py)
• Data/storage: CSV, NumPy (.npy), Pickle (.pkl)
• Visualization: CGI-Python backend, HTML5/CSS3/JavaScript frontend
• Environment management: Conda (environment.yml in each module)