neuralset.events.transforms.chunking.ChunkEvents

class neuralset.events.transforms.chunking.ChunkEvents(*, infra: Backend | None = None, event_type_to_chunk: Literal['Audio', 'Video'], event_type_to_use: str | None = None, min_duration: float | None = None, max_duration: float = inf)

This functions chunks long events (audio or video) into shorter events. It supports two modes:

1. Duration-based chunking (if event_type_to_use is None):

   • Each event of type event_type_to_chunk is split into consecutive chunks with duration between min_duration and max_duration.

   • Ensures that no chunk exceeds max_duration.

This avoids OOM errors when feeding long events into a deep learning model (e.g. Wav2Vec).

Example:

input:
    min_duration: 2
    max_duration: 3
    events:
        sound:    [.......]
out:
    events:
        sound1:   [...]
        sound2:      [....]

2. Split-based chunking (if event_type_to_use is not None):

   • Events of type event_type_to_chunk are split according to the train/val/test splits of event_type_to_use.

   • Ensures each chunk respects min_duration and max_duration.

   • Prevents data leakage when processing long events with a deep learning model (e.g. Wav2Vec.

   • Requires that the splits for event_type_to_use are already assigned.

   Example:

   input:
       max_duration: 2
       event_type_to_use: Word
       events:
           sound:    [.......]
           word :    [1112233]
   out:
       events:
           sound1:   [..]
           sound2:     [.]
           sound3:      [..]
           sound4:        [..]