How sigcgans/conditional-sig-wasserstein-gans works
Project Type and Purpose
• This is a research‐oriented PyTorch codebase implementing Conditional Sig-Wasserstein GANs (SigCWGAN) for time-series data generation, with scripts for training, evaluation, and baseline comparisons.
File Structure Analysis
• Top-Level Scripts
• train.py – orchestrates model training (loads data, configures hyperparameters, instantiates models, logs results)
• evaluate.py – runs evaluation metrics on saved checkpoints
• hyperparameters.py – central location for default training/evaluation settings
• Library Code (lib/)
• lib/data.py – data loading, dataset wrappers (synthetic VAR/ARCH and real stock data)
• lib/augmentations.py – time-series augmentation utilities
• lib/algos/ – GAN implementations
– base.py (abstract GAN interface)
– gans.py (standard GAN)
– gmmn.py (GAN with MMD)
– sigcwgan.py (the Sig-Wasserstein variant)
• lib/arfnn.py – autoregressive feedforward neural network baseline
• lib/test_metrics.py – evaluation metrics module
• lib/plot.py – visualization helpers
• lib/utils.py – shared utility functions (logging, checkpointing)
• Config & Environment
• requirements.yml – conda environment spec
• .gitignore
Key Architectural Patterns
• Modular, plug-and-play algorithm implementations via a common base class (lib/algos/base.py)
• Separation of concerns: data ingestion → model definitions → training loop → evaluation → plotting/storage
• Script-driven orchestration (train.py and evaluate.py) rather than a web API
a. Data Layer
• Synthetic Data Generator (VAR, ARCH)
• Real-world Data Loader (stock data)
• Augmentation Pipeline
b. Model Layer (lib/algos)
• SigCWGAN core (generator, discriminator, scoring)
• Baseline GANs (RCGAN, TimeGAN, GMMN)
• ARFNN baseline
c. Orchestration Layer
• Training Script (train.py)
• Evaluation Script (evaluate.py)
• Hyperparameter Manager (hyperparameters.py)
d. Utility Layer
• Logging & Checkpointing (utils.py)
• Plotting & Metrics (plot.py, test_metrics.py)
e. Results Storage
• numerical_results/ (training logs, CSV summaries, model checkpoints)
Component Interactions & Data Flow
1. train.py reads hyperparameters → calls data.py → obtains data batches
2. Data batches pass through augmentations.py → fed into selected lib/algos model
3. Model’s forward/backward loops handled by train.py → util.py manages checkpoints & logging
4. At intervals, train.py invokes plot.py to generate figures → stored under numerical_results
5. evaluate.py loads saved checkpoints → uses test_metrics.py to compute metrics → writes summary.csv