How arberetum/synthetic_eeg works
Main Components of the System
• Data Input & Preprocessing Module:
• Raw EEG data source (from the Kaggle Seizure Prediction Challenge dataset) as the starting point.
• "preprocessing.ipynb" notebook that handles initial data cleaning and preparation.
• Feature Extraction and Selection:
• A component for "feature_extraction.py" and "feature_selection.py". These scripts are responsible for extracting relevant signal features from the preprocessed data.
• Generative Model Training:
• WGAN: Represented by the "WGAN.ipynb" notebook. Emphasize that this module uses an adversarial framework with a critic network and employs the earth-mover distance to ensure training stability.
• DDPM: Represent the DDPM-based components which are spread across "DDPM-Interictal.ipynb", "DDPM-Preictal.ipynb", and "DDPM-smaller_data.ipynb". These notebooks indicate iterative denoising refinement of synthetic signals.
• DCGAN if it is used as an auxiliary experiment.
• Evaluation & Analysis Module:
• "figures.ipynb", where loss curves and other performance metrics (plus qualitative visualizations like spectrogram comparisons) are generated.
• Reporting & Documentation:
• Incorporate the CS_534_Final_Report_Shoibolina_Aiden.pdf and README as final artifacts that summarize the methods, experiments, and outcomes.
Relationships and Interactions
• From raw EEG data to the Preprocessing Module.
• From preprocessing to Feature Extraction/Selection.
• From there, split the flow to the generative model training pipelines (WGAN and DDPM).
• After model training, the generated synthetic data flows to the Evaluation Module where classification models compare performance with real data.
• Demonstrate feedback loops if applicable (e.g., evaluation results influencing further model tuning).
Highlighting Architectural Patterns
• Data Pipeline: Emphasize the sequential processing of data from raw input to preprocessed signals, then through feature extraction and model training.
• Experimental Separation: Clearly distinguish between the different generative modeling approaches (WGAN vs. DDPM), showing that they operate in parallel to generate synthetic data.
• Modular Design: Indicate that each notebook/script serves as an independent module that can be modified or replaced without affecting the overall pipeline.
Essential Elements
• Clear labels for each module:
• “Raw EEG Data Input” → “Preprocessing (preprocessing.ipynb)” → “Feature Extraction/Selection (feature_extraction.py & feature_selection.py)” → Branches into “WGAN (WGAN.ipynb)” and “DDPM Models (DDPM-*.ipynb)” → “Evaluation (evaluation_models.py, figures.ipynb)” → “Reporting (PDF report & README)”
• Data/Artifact components (raw data, synthetic data, output figures)
• Processing modules (preprocessing, feature extraction, model training, evaluation)
• Documentation elements (final report, README)