How jayoungkim408/stasy works
Project Type and Purpose
• This is a research-oriented Python project that implements a score-based generative model for tabular data synthesis.
• It provides training, fine-tuning, sampling, and evaluation workflows driven by a single entry point (main.py) and configurable via Python config scripts in configs/.
Main Architectural Layers
A. Configuration & CLI Layer
• main.py
• configs/ directory (e.g., configs/shoppers.py, configs/iris.py)
• CLI arguments (--config, --mode, --workdir)
B. Data Layer
• Raw tabular datasets under tabular_datasets/ (JSON/NPZ)
• Dataset loader modules (datasets.py, datasets_tabular.py, prepare_dataset.py, prepare_dataset_utils.py)
• Preprocessing utilities in utils.py, tabular_utils.py
C. Model Layer
• Core score-based model implementation in models/:
– SDE solvers (sde_lib.py)
– Neural net modules (models/layers.py, models/layerspp.py)
– Model definitions (models/ncsnpp_tabular.py)
– EMA wrapper (models/ema.py)
– Misc utilities (models/utils.py, models/tabular_utils.py)
D. Training/Sampling/Evaluation Engine
• Training loop & loss computation (losses.py, likelihood.py, run_lib.py, sampling.py, evaluation.py)
• Checkpointing & logging (via run_lib.py or custom code in main.py)
E. External Dependencies & Environments
• Python libraries (PyTorch, NumPy, SciPy, Hydra or custom config loader)
• Conda environment (requirements.yaml)
• GPU/Compute cluster (optional)
Relationships and Data Flows
• CLI → Config Loader: Parse arguments & load Python config
• Config → Data Layer: Select dataset paths & preprocessing parameters
• Data Loader → Preprocessing → Model Input
• Model Layer ↔ SDE Module: Forward/backward passes, score estimation
• Training Engine → Loss Module → Optimizer → Model Layer (iterative)
• Model Layer → EMA Module → Checkpoint Manager
• Sampling Engine (invoked by main.py --mode eval) → SDE Module → Generated Samples
• Evaluation Module → Metrics & Results → Output directory under workdir
Architectural Patterns & Principles
• Layered separation: Config/CLI, Data, Model, Training/Evaluation
• Modular design: Each concern (data prep, model, sampling, evaluation) lives in its own module
• Reusable utilities: Shared functions in utils.py, tabular_utils.py
• Experiment reproducibility: Checkpoints and logging in a per-workdir folder