boaztulu/EthioDance-ViT

PyTorch implementation and benchmarks for spatiotemporal classification of Ethiopian traditional dances using Video Vision Transformers. # EthioDance-ViT PyTorch implementation and benchmarks for **spatiotemporal classification of Ethiopian traditional dances** using Video Vision Transformers (TimeSformer). --- ## Dataset 1,806 video clips across 5 Ethiopian dance styles: | Class | Clips | | -------- | ----- | | Amhara | 516 | | Woliyta | 337 | | Oromo | 334 | | Gurageya | 316 | | Tigriga | 303 | The raw dataset is stored **outside** the repository (it is too large for git). By default the code expects it at `../Data/ALL/ /*.mp4` relative to this repo; override with `data.root` in the YAML config. ## Layout ``` EthioDance-ViT/ # / # /best.pth ``` ## HiPerGator ```bash sbatch scripts/train_hipergator.sh # full run, requeue-safe sbatch scripts/train_hipergator.sh --ablation ``` ## Class-imbalance strategy We combine three complementary techniques (all toggleable in YAML): 1. **Class-Balanced Focal Loss** (Cui et al., CVPR 2019) — effective-number reweighting with a focusing term that down-weights easy majority examples. 2. **Weighted random sampler** — produces batches that are class-balanced in expectation, so every batch trains every class. 3. **Video MixUp** — clip-level linear interpolation to regularize the overrepresented Amhara class without discarding data. ## Citation _(to be added upon publication)_