avalanche.benchmarks.split_continuous_linear_decay_stream

avalanche.benchmarks.split_continuous_linear_decay_stream(original_stream: Iterable[DatasetExperience[TCLDataset]], experience_size: int, iters_per_virtual_epoch: int, beta: float, shuffle: bool) → CLStream[DatasetExperience[TCLDataset]][source]
Creates a stream of sub-experiences from a list of overlapped

experiences with a linear decay in the overlapping areas.

Parameters:

  • original_stream – The original stream.

  • experience_size – The size of each sub-experience.

  • iters_per_virtual_epoch – The number of iterations per virtual epoch. This parameter determines the number of (sub-)experiences that we want to create from each experience in the original stream, after “merging” all experiences with a certain level of “overlap”.

  • beta – The beta parameter for the linear decay function which indicates the amount of overlap.

  • shuffle – Whether to shuffle the sub-experiences.

Terminology is taken from the official implementation of the paper: “Task Agnostic Continual Learning Using Online Variational Bayes” by Zero et al. (https://arxiv.org/abs/2006.05990) Code repo: https://github.com/chenzeno/FOO-VB/tree/main

Returns:

A stream of sub-experiences.