How seanzhang99/superhugeaad works
SuperHugeAAD is a framework for auditory-attention decoding from preprocessed EEG: it supports envelope reconstruction and spatial-attention classification across multiple datasets. The main workflow configures a task and cross-validation run, selects and splits metadata, loads EEG/audio samples and applies transforms, then trains and evaluates a selected model. The README documents model families, cross-validation options, and MATLAB interoperability; some of those implementation details were not sampled, so their internal connections are not asserted here.
Main entry points: data_interface.py Data module · main.py DNN entry point · model_interface.py Model interface · kulavgc.py Linear entry point
Components 26
Experiment orchestration
- DNN entry point
train/main.py - Linear entry point
train/kulavgc.py - Multi-run CLI
superhuge/utils/multi_run_cli.py - Task and CV parser
superhuge/utils/task_config_parser.py
Data pipeline
- Trial metadata
superhuge/data/metadata_processing/data.py - Metadata filters
superhuge/data/metadata_filters/abc.py - Classification filters
superhuge/data/metadata_filters/classify_filter.py - Regression filters
superhuge/data/metadata_filters/regress_filter.py - Filter composer
superhuge/data/metadata_filters/composer.py - Cross-validation grouping
superhuge/data/metadata_processing/group.py - Data module
superhuge/data/interface/data_interface.py - EEG sample dataset
superhuge/data/datasets/eeg_dataset.py - Regression dataset
superhuge/data/datasets/eeg_regression_base_dataset.py - Classification dataset
superhuge/data/datasets/eeg_classify_base_dataset.py - Transform API
superhuge/data/transforms/abc.py - Transform pipeline
superhuge/data/transforms/composer.py - EEG transforms
superhuge/data/transforms - MATLAB interoperability
superhuge/data/interface/matlab_utils.py
AAD models
- Model interface
superhuge/model/interface/model_interface.py - Deep architectures
superhuge/model - Linear baselines
superhuge/model/linear
Model support
- Task losses and metrics
superhuge/model/loss - Neural building blocks
superhuge/model/module - Channel mapping
superhuge/model/interface/channel_mapping_interface.py
Other components
- Experiment user
- Preprocessed EEG data
Connections 26
- Experiment user → DNN entry point: starts DNN run
README.md - Experiment user → Linear entry point: starts linear run
README.md - DNN entry point → Multi-run CLI: invokes
train/main.py - Linear entry point → Multi-run CLI: imports
train/kulavgc.py - Multi-run CLI → Task and CV parser: uses task config
superhuge/utils/multi_run_cli.py - Preprocessed EEG data → Data module: loads preprocessed files
README.md - Data module → Trial metadata: uses metadata
superhuge/data/interface/data_interface.py - Data module → Metadata filters: configures filters
superhuge/data/interface/data_interface.py - Data module → Classification filters: checks filter type
superhuge/data/interface/data_interface.py - Data module → Regression filters: imports filter
superhuge/data/interface/data_interface.py - Data module → Filter composer: configures filters
superhuge/data/interface/data_interface.py - Data module → Cross-validation grouping: uses grouping
superhuge/data/interface/data_interface.py - Data module → EEG sample dataset: creates datasets
superhuge/data/interface/data_interface.py - Data module → Regression dataset: supports dataset class
superhuge/data/interface/data_interface.py - Data module → Classification dataset: supports dataset class
superhuge/data/interface/data_interface.py - EEG sample dataset → Trial metadata: reads metadata
superhuge/data/datasets/eeg_dataset.py - EEG sample dataset → Transform pipeline: applies transforms
superhuge/data/datasets/eeg_dataset.py - Transform pipeline → Transform API: composes transforms
superhuge/data/transforms/composer.py - Transform pipeline → EEG transforms: runs transform types
superhuge/data/transforms/composer.py - Model interface → Neural building blocks: uses components
superhuge/model/interface/model_interface.py - DNN entry point → Data module: trains with data module
- DNN entry point → Model interface: trains model interface
- Model interface → Deep architectures: instantiates selected model
superhuge/model/interface/model_interface.py - Model interface → Task losses and metrics: uses configured loss
- Linear entry point → Linear baselines: runs linear baseline
- Data module → MATLAB interoperability: can be created via MATLAB
README.md