peal.data.dataloaders.DataStack

class peal.data.dataloaders.DataStack

Bases: object

Per-class FIFO buffer of samples drawn from a dataset or dataloader.

data[c] holds [X, y] pairs of class c. The stack is refilled from the source whenever some class runs empty, so pop(c) always returns a sample of class c; it is used to assemble class-ordered batches for visualizations.

Parameters:

• datasource (torch.utils.data.Dataset or DataLoader or DataloaderMixer) – Where samples come from. A dataset is walked sequentially and cyclically via current_idx; a dataloader is drawn from with next(iter(...)).

• num_classes (int) – Number of classes, i.e. number of per-class lists.

• transform (callable, optional) – Transform temporarily swapped into dataset.transform while the stack is being filled.

Methods

fill_stack	Draw samples from the source until every class list is non-empty.
pop	Remove and return the oldest buffered sample of one class.
reset	Empty all class lists, reset a DataloaderMixer source and refill.

__init__(datasource, num_classes, transform=None)

Create the per-class lists and fill them once.

fill_stack()

Draw samples from the source until every class list is non-empty.

With hints or indices enabled the label is a tuple and its first element is used as the class index.

pop(class_idx)

Remove and return the oldest buffered sample of one class.

Parameters:

class_idx (int) – Class whose sample is popped.

Returns:

[X, y] of the sample; the stack is refilled afterwards.

Return type:

list

reset()

Empty all class lists, reset a DataloaderMixer source and refill.