A refactored version, with cleaner and more understandable internal code.
2.1
Fixed previous bug. Added support for multiple integer labels per clip.
2.0
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
1.0
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/