Data¶

SONNDataset pairs features with targets for a PyTorch DataLoader; the preprocessing module holds the helpers for splitting data by row position. See Data.

SONNDataset ¶

SONNDataset(x: ArrayLike, target: ArrayLike | None, split: int | None = None)

Bases: Dataset

Map-style dataset of (x, target) rows, optionally one half of them.

Parameters:

Name	Type	Description	Default
`x`	`(N, d) array or tensor`	Input rows.	required
`target`	`(N, ...) array or tensor, or None`	Targets. None yields `(x, None)` items, which PyTorch's default collate function cannot batch; predict unlabeled rows with `SONN.infer` on a tensor instead.	required
`split`	`(None, 0, 1)`	None uses every row; 0 the even rows (0, 2, 4, ...); 1 the odd rows. The trainer uses the two halves as subsets A and B of one loader.	`None`

preprocessing ¶

DataSetType ¶

Bases: IntEnum

Tag attached to each row of a dataset by set_split_types.

SequenceTypeSet ¶

Bases: IntEnum

How rows are assigned to train vs. validate.

sqRandom — random Bernoulli per row. sqMode1 — every 2nd row is validate. sqMode3_1 — every 3rd row is validate. sqMode4_1 — every 4th row is validate. sqMode2 — every 2nd row is train (inverse of sqMode1). sqMode3_2 — every 3rd row is train (inverse of sqMode3_1). sqMode4_2 — every 4th row is train (inverse of sqMode4_1).

is_mode1_type `classmethod` ¶

is_mode1_type(seq_type: SequenceTypeSet) -> bool

True for the modes that put every k-th row in validate (sqMode1, sqMode3_1, sqMode4_1).

is_mode2_type `classmethod` ¶

is_mode2_type(seq_type: SequenceTypeSet) -> bool

True for the modes that put every k-th row in train (sqMode2, sqMode3_2, sqMode4_2).

train_preprocessing ¶

train_preprocessing(data_x: ndarray | DataFrame, data_y: ndarray | DataFrame | Series, feature_names: Sequence[str] | None) -> tuple[np.ndarray, np.ndarray]

process of train input data: transform to numpy matrix, transpose etc

predict_preprocessing ¶

predict_preprocessing(data_x: ndarray | DataFrame, n_features: int) -> tuple[np.ndarray, int]

process of predict input data: transform to numpy matrix, transpose etc

set_split_types ¶

set_split_types(seq_type: SequenceTypeSet, data_len: int) -> np.ndarray

Set seq_types array that will be used to divide data set to train and validate ones

split_dataset ¶

split_dataset(data_x: ndarray, data_y: ndarray, seq_type: SequenceTypeSet) -> tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]

Split train and validate data sets from input data set and target