ml.tasks.datasets.clippify

Defines a dataset for converting frame streams to clips.

class ml.tasks.datasets.clippify.ClippifyDataset(image_dataset: IterableDataset[Batch], num_images: int, *, stride: int = 1, jump_size: int = 1, sample_first: bool = False, use_last: bool = False)[source]

Bases: IterableDataset[tuple[Tensor, Batch]], Generic[Batch]

Defines a dataset which efficiently yields sequences of images.

The underlying image dataset just needs to iterate through a sequence of images, in order. This wrapper dataset collates the images into clips with some striding between adjacent images.

Images are inserted into a deque and routinely popped. The underlying dataset should do necessary error handling, since this dataset will simply throw an error on failure.

Parameters:

  • image_dataset – The child dataset which yields the images

  • num_images – The number of images in each clip

  • stride – The stride between adjacent images

  • jump_size – How many frames to jump in the future

  • sample_first – If set, don’t always start on the first item; instead, sample the first item within jump_size

  • use_last – If set, always use the last item in the dataset

image_iter: Iterator[Batch]
inds: list[int]
image_queue: Deque[tuple[int, Batch]]
image_ptr: int
image_queue_ptr: int
hit_last: bool