Configuration

Dataclass configuration objects used across the package.

Classes:

Name	Description
DatasetPaths	Filesystem locations used by data and experiment workflows.
PreprocessingConfig	Preprocessing and augmentation settings.
DataSplitConfig	Train/validation/test split settings.
TrainingConfig	Training defaults shared by examples and notebook runs.
ModelConfig	Model architecture defaults.

DatasetPaths dataclass

DatasetPaths(
    raw_dir: Path = Path("data/raw"),
    processed_dir: Path = Path("data/processed"),
    local_cache_dir: Path | None = Path("data_temp"),
    artifacts_dir: Path = Path("artifacts/runs"),
)

Filesystem locations used by data and experiment workflows.

Attributes:

Name	Type	Description
raw_dir	Path	Directory for official downloaded BCI IV-2a archives and extracted GDF files.
processed_dir	Path	Directory for generated project .npy arrays.
local_cache_dir	Path | None	Optional directory containing an existing local cache of the six .npy files.
artifacts_dir	Path	Directory for generated reports, metrics, and plots.

PreprocessingConfig dataclass

PreprocessingConfig(
    sub_sample: int = 2,
    average: int = 2,
    noise: bool = True,
    noise_std: float = 0.5,
    max_time_step: int = 500,
    n_classes: int = 4,
    label_offset: int = 769,
    seed: int = 1,
)

Preprocessing and augmentation settings.

Attributes:

Name	Type	Description
sub_sample	int	Temporal stride used for max-pooling and subsampling augmentation.
average	int	Window size used for average-pooling augmentation.
noise	bool	Whether to add Gaussian noise to averaged/subsampled augmentations.
noise_std	float	Standard deviation of Gaussian augmentation noise.
max_time_step	int	Number of raw time samples kept before temporal reduction.
n_classes	int	Number of output classes.
label_offset	int	Offset used to convert BCI cue labels 769-772 to 0-3.
seed	int	Random seed for deterministic validation splits and augmentation.

DataSplitConfig dataclass

DataSplitConfig(
    validation_ratio: float = 0.17,
    subject_id: int | None = None,
    seed: int = 1,
)

Train/validation/test split settings.

Attributes:

Name	Type	Description
validation_ratio	float	Fraction of train/valid trials assigned to validation.
subject_id	int | None	Optional subject ID 0-8; None uses all subjects.
seed	int	Random seed used for validation-index sampling.

TrainingConfig dataclass

TrainingConfig(
    epochs: int = 100,
    gan_epochs: int = 20,
    batch_size: int = 64,
    gan_batch_size: int = 128,
    learning_rate: float = 0.001,
    gan_learning_rate: float = 0.0003,
    use_gan_augmentation: bool = False,
    gan_samples_per_class: int = 10,
    seed: int = 1,
    fast_dev_run: bool = False,
)

Training defaults shared by examples and notebook runs.

Attributes:

Name	Type	Description
epochs	int	Number of classifier training epochs for full reproduction.
gan_epochs	int	Number of conditional GAN training epochs.
batch_size	int	Classifier batch size.
gan_batch_size	int	Conditional GAN batch size.
learning_rate	float	Classifier Adam learning rate.
gan_learning_rate	float	Conditional GAN Adam learning rate.
use_gan_augmentation	bool	Whether to append synthetic GAN samples to the training split.
gan_samples_per_class	int	Number of interpolation samples generated for each ordered class pair in the GAN augmentation routine.
seed	int	Framework and data random seed.
fast_dev_run	bool	If True, notebooks/trainers run a one-epoch smoke check.

ModelConfig dataclass

ModelConfig(
    channels: int = 22,
    n_classes: int = 4,
    max_time_step: int = 500,
    dropout: float = 0.5,
    cnn_filters: tuple[int, int, int, int] = (
        25,
        50,
        100,
        200,
    ),
    cnn_kernel: tuple[int, int] = (10, 1),
    lstm_units: tuple[int, int, int] = (32, 16, 8),
    cnn_lstm_units: int = 10,
    latent_dim: int = 128,
)

Model architecture defaults.

Attributes:

Name	Type	Description
channels	int	Number of EEG channels.
n_classes	int	Number of classification targets.
max_time_step	int	Raw time samples used before preprocessing reduction.
dropout	float	Dropout probability used in CNN blocks.
cnn_filters	tuple[int, int, int, int]	Filters for the four convolutional blocks.
cnn_kernel	tuple[int, int]	Convolution kernel size for CNN-style classifiers.
lstm_units	tuple[int, int, int]	Hidden sizes for the stacked bidirectional LSTM classifier.
cnn_lstm_units	int	Hidden size for the CNN-LSTM recurrent layer.
latent_dim	int	Latent noise dimension used by conditional GAN components.