lisbet.config.schemas

Classes

DataAugmentationConfig(**data)	Configuration for a single data augmentation technique.
DataAugmentationPipeline(**data)
DataConfig(**data)
ExperimentConfig(**data)
LSTMBackboneConfig(**data)
ModelConfig(**data)
TCNBackboneConfig(**data)
TrainingConfig(**data)
TransformerBackboneConfig(**data)

class lisbet.config.schemas.TransformerBackboneConfig(**data)

type: Literal['transformer']

feature_dim: int | None

embedding_dim: int

hidden_dim: int

num_heads: int

num_layers: int

max_length: int

model_config: ClassVar[ConfigDict] = {}

Configuration for the model, should be a dictionary conforming to [ConfigDict][pydantic.config.ConfigDict].

class lisbet.config.schemas.TCNBackboneConfig(**data)

type: Literal['tcn']

feature_dim: int | None

embedding_dim: int

hidden_dim: int

num_layers: int

kernel_size: int

dilation_base: int

dropout: float

model_config: ClassVar[ConfigDict] = {}

Configuration for the model, should be a dictionary conforming to [ConfigDict][pydantic.config.ConfigDict].

class lisbet.config.schemas.LSTMBackboneConfig(**data)

type: Literal['lstm']

feature_dim: int | None

embedding_dim: int

hidden_dim: int

num_layers: int

model_config: ClassVar[ConfigDict] = {}

Configuration for the model, should be a dictionary conforming to [ConfigDict][pydantic.config.ConfigDict].

class lisbet.config.schemas.DataConfig(**data)

data_path: str

data_format: str

data_scale: str | None

data_filter: str | None

select_coords: str | None

rename_coords: str | None

window_size: int

window_offset: int

fps_scaling: float

dev_ratio: float | None

train_sample: float | None

dev_sample: float | None

model_config: ClassVar[ConfigDict] = {}

Configuration for the model, should be a dictionary conforming to [ConfigDict][pydantic.config.ConfigDict].

class lisbet.config.schemas.DataAugmentationConfig(**data)

Configuration for a single data augmentation technique.

Augmentation families and parameter semantics:

Permutation-based :

• all_perm_id: Full-window permutation of individual identities.

• all_perm_ax: Full-window permutation of spatial axes.

• blk_perm_id: Block (contiguous frames) permutation of individual

  identities. Uses frac for relative block length.

For these, p is the probability of applying the entire transform (implemented via RandomApply in the pipeline).

Jitter-based :

• gauss_jitter: For the full window (time, keypoints,

  individuals), adds N(0, sigma) noise.

Ablation-based :

• kp_ablation: Per-element Bernoulli(pB) mask over (time, keypoints,

  individuals), sets selected elements to 0.0 (all space dims). Simulates missing or occluded keypoints.

Rotation-based :

• rotation: Random rotation of all keypoint coordinates around the

  center of the normalized coordinate space. Supports 2D and 3D (auto-detected). Uses max_angle for the rotation range and mode for post-rotation normalization.

Attributes:

name: Name of the augmentation technique p: Probability of applying this transformation (0.0 to 1.0) pB: When applicable, per-element Bernoulli probability (kp_ablation types only) frac: Fraction of frames to permute (only for block-based augmentations, 0.0

to 1.0 exclusive)

sigma: Standard deviation of Gaussian noise (jitter types only) max_angle: Maximum rotation angle in degrees (rotation only, default 180.0) mode: Post-rotation normalization mode (rotation only). One of “truncate”,

“rescale”, “none”. Default “truncate”.

model_config: ClassVar[ConfigDict] = {'extra': 'forbid'}

Configuration for the model, should be a dictionary conforming to [ConfigDict][pydantic.config.ConfigDict].

name: Literal['all_perm_id', 'all_perm_ax', 'blk_perm_id', 'gauss_jitter', 'kp_ablation', 'rotation']

p: float

pB: float | None

frac: float | None

sigma: float | None

max_angle: float | None

mode: str | None

VALID_PARAMS: ClassVar[dict[str, set[str]]] = {'all_perm_ax': {'p'}, 'all_perm_id': {'p'}, 'blk_perm_id': {'frac', 'p'}, 'gauss_jitter': {'p', 'sigma'}, 'kp_ablation': {'p', 'pB'}, 'rotation': {'max_angle', 'mode', 'p'}}

classmethod validate_probability(v)

classmethod validate_fraction(v, info)

classmethod validate_sigma(v)

classmethod validate_pB(v)

classmethod validate_max_angle(v)

classmethod validate_mode(v)

validate_parameters_for_augmentation()

Ensure only valid parameters are set for each augmentation type and apply defaults.

class lisbet.config.schemas.DataAugmentationPipeline(**data)

augmentations: list[DataAugmentationConfig]

model_config: ClassVar[ConfigDict] = {}

Configuration for the model, should be a dictionary conforming to [ConfigDict][pydantic.config.ConfigDict].

class lisbet.config.schemas.ModelConfig(**data)

model_id: str | None

backbone: Annotated[TransformerBackboneConfig | LSTMBackboneConfig | TCNBackboneConfig, FieldInfo(annotation=NoneType, required=True, discriminator='type')]

out_heads: dict[str, dict]

input_features: dict[str, list[str]]

window_size: int

window_offset: int

model_config: ClassVar[ConfigDict] = {}

Configuration for the model, should be a dictionary conforming to [ConfigDict][pydantic.config.ConfigDict].

class lisbet.config.schemas.TrainingConfig(**data)

epochs: int

batch_size: int

learning_rate: float

data_augmentation: list[DataAugmentationConfig] | None

save_weights: str | None

save_history: bool

mixed_precision: bool

head_type: Literal['mlp', 'linear']

freeze_backbone_weights: bool

load_backbone_weights: str | Path | None

model_config: ClassVar[ConfigDict] = {}

Configuration for the model, should be a dictionary conforming to [ConfigDict][pydantic.config.ConfigDict].

class lisbet.config.schemas.ExperimentConfig(**data)

run_id: str | None

model: ModelConfig

training: TrainingConfig

data: DataConfig

task_ids_list: list[str]

task_data: str | None

seed: int

output_path: Path

model_config: ClassVar[ConfigDict] = {}

Configuration for the model, should be a dictionary conforming to [ConfigDict][pydantic.config.ConfigDict].