Struct DatasetPlan 

pub struct DatasetPlan { /* private fields */ }

Immutable training-data plan. Cheap to clone (TidyView holds Rc<DataFrame>).

Implementations

impl DatasetPlan

pub fn from_view(source: TidyView) -> Self

pub fn from_dataframe(df: DataFrame) -> Self

pub fn with_features(self, cols: Vec<String>) -> Self

pub fn with_label(self, col: String) -> Self

pub fn with_encoding(self, col: String, enc: EncodingSpec) -> Self

pub fn with_split(self, split: SplitSpec) -> Self

pub fn with_batch(self, batch: BatchSpec) -> Self

pub fn nrows(&self) -> usize

pub fn n_features(&self) -> usize

pub fn feature_cols(&self) -> &[String]

pub fn label_col(&self) -> Option<&str>

pub fn split_spec(&self) -> &SplitSpec

pub fn batch_spec(&self) -> &BatchSpec

pub fn plan_hash(&self) -> Option<&[u8; 32]>

pub fn validate(&self) -> Result<(), DatasetError>

Validate the plan against the source schema. Cheap; no materialization. Called automatically by iter_batches and split_rows; useful in tests / dry-runs.

pub fn split_rows(&self, which: Split) -> Result<Vec<u32>, DatasetError>

Ascending row IDs assigned to which. Row IDs are indices into the materialized source (post-filter, post-select), not the raw underlying DataFrame.

pub fn iter_batches(&self, which: Split) -> Result<BatchIterator, DatasetError>

Iterate batches over which split. Each batch is fully resolved into row-major Tensors. Categorical dictionaries are built over the entire materialized source (so val/test see codes consistent with train) and frozen before iteration begins.

Trait Implementations

impl Clone for DatasetPlan

fn clone(&self) -> DatasetPlan

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.