nexuml.data.dataset¶
Base dataset contracts and metadata-backed helpers for NexuML.
Bases: Dataset
Base dataset returning (x, y) TensorDict pairs.
The base class supports two common patterns used throughout NexuML:
- metadata-backed datasets via
self.meta - in-memory datasets via
self.data
Subclasses can still override __getitem__ entirely for bespoke behavior.
split_meta(split_ratio: list[float]) -> None
Convert 'fit' rows to train/val and 'all' rows to train/val/test.
get_split(split: str) -> Self
Return a view of this dataset filtered to a single split.
Returns:
| Type | Description |
|---|---|
Self
|
A new :class: |
Raises:
| Type | Description |
|---|---|
ValueError
|
If the dataset has no metadata. |
load_item(idx: int, row: Series) -> TensorDict
Load features for a single sample.
Subclasses can override this to lazily load file-backed content.
Returns:
| Type | Description |
|---|---|
TensorDict
|
Feature tensor dictionary for the sample. |
load_labels(idx: int, row: Series) -> TensorDict | None
Load labels for a single sample.
By default labels are read from self.meta using self.label_names.
Subclasses can override this for file-backed or computed targets.
Returns:
| Type | Description |
|---|---|
TensorDict | None
|
Label tensor dictionary, or |