v2.2

A refactored version, with cleaner and more understandable internal code.

Stable

Fixed previous bug.
Added support for multiple integer labels per clip.

Buggy Release

Breaking change from 1.0, where in annotations.txt you now need to supply START_FRAME and END_FRAME per sample instead of just NUM_FRAMES. This allows you to use small clips from a whole video if needed.

BUG:
Didn't load enough frames when FRAMES_PER_SEGMENT > 1

Stable

Load videos into PyTorch
Features:

• A frame sampling strategy that helps with training (Sparse Temporal Sampling - "Temporal Segment Networks")
• A very fast video loading pipeline that eliminates input bottlenecks
• Generic and therefore usable with custom datasets with minimum effort
• Easy out-of-the-box support for torchvision 0.8.0's video-level augmentations/transforms (identical image transforms on an image batch)

Extensive documentation at https://video-dataset-loading-pytorch.readthedocs.io/