How Do You Handle Variable-Length Video Clips in a PyTorch DataLoader?
PyTorch Forums
How Do You Handle Variable-Length Video Clips in a PyTorch DataLoader?
Category: vision I am building a video training pipeline where samples have different numbers of frames and resolutions. For each sample, I currently have: - Video tensor with shape [T, C, H, W] - Frame-level labels - Optional timestamps - Different values of T, H, and W I am comparing three approaches: 1. Sample a fixed temporal window and resize frames before batching 2. Pad clips to the longest sequence in each batch and return a mask 3. Use length-based bucketing to reduce padding overh...
0 comments
No comments yet.