progressive resizing

**Progressive Resizing** is a **training technique that starts training with small, low-resolution images and progressively increases the resolution** — inspired by progressive growing in GANs, this approach yields faster training and often better generalization by building feature hierarchies from coarse to fine. **How Progressive Resizing Works** - **Start Small**: Begin training with small images (e.g., 64×64) — fast iterations, rapid feature learning. - **Increase**: Periodically double the resolution (64→128→224→448) — model refines features at each scale. - **Learning Rate**: Optionally reset or warm up the learning rate at each resolution increase. - **Transfer**: Lower-resolution features transfer to higher resolution — warm-starting accelerates training. **Why It Matters** - **Speed**: Low-resolution training is 4-16× faster — majority of training epochs run at low resolution. - **Regularization**: Starting at low resolution acts as a regularizer — model learns to extract the most important features first. - **fast.ai**: Popularized by fast.ai as a key technique for efficient, high-quality training. **Progressive Resizing** is **training from blurry to sharp** — starting with fast low-resolution training and progressively refining to full resolution.