Skip to content

Layout#

Layout instances are responsible for splitting your datasets and managing which data fold is used for each experiment. They also check reproducibility of your data splits.

Usage#

Layout is a cornerstone of every built experiment. Via layout one can define structure of their data splits.

After building an experiment one may want to extract data of created splits. Heretofore it was done by manually loading the ids. Thunder's Layouts provides interface for extracting them by simply loading experiment config. 

import lazycon
cfg = lazycon.load("/path/to/experiment.config")
cfg.layout.set(fold=0)
train_ids = cfg.layout.train # cfg.layout.SPLIT_NAME

Thunder Layouts#

Name Description
Split Layout for K fold cross-validation
SingleSplit Layout with several sets (e.g. train + val + test)
FixedSplit Creates layout from predefined K-fold split
FixedSingleSplit Creates single fold layout from predefined data split.

All Layout subclasses follow common interface

thunder.layout.interface.Layout #

Bases: ABC

Source code in thunder/layout/interface.py 
21
22
23
24
25
26
27
28
29
30
31
32
class Layout(ABC):
    @abstractmethod
    def build(self, experiment: Path, config: Config) -> Iterable[Node]:
        pass

    @abstractmethod
    def load(self, experiment: Path, node: Optional[Node]) -> Tuple[Config, Path, Dict[str, Any]]:
        pass

    @abstractmethod
    def set(self, **kwargs):
        pass