multi-resolution training

**Multi-Resolution Training** is a **training strategy that exposes the model to inputs at multiple spatial resolutions during training** — enabling the model to learn features at different scales and perform well regardless of the input resolution encountered at inference time. **Multi-Resolution Methods** - **Random Resize**: Randomly resize training images to different resolutions within a range each iteration. - **Multi-Scale Data Augmentation**: Apply scale augmentation as part of the data augmentation pipeline. - **Resolution Schedules**: Train at low resolution first, progressively increase to high resolution. - **Multi-Branch**: Process multiple resolutions simultaneously through parallel branches. **Why It Matters** - **Robustness**: Models trained at a single resolution often fail when tested at different resolutions. - **Efficiency**: Lower-resolution training is faster — multi-resolution training can start fast and refine. - **Deployment**: Edge devices may need different resolutions — multi-resolution training prepares one model for all. **Multi-Resolution Training** is **learning at every zoom level** — training models to handle any input resolution by exposing them to multiple scales during training.