gridmix

**GridMix** is a **data augmentation technique that divides images into a grid and randomly assigns each cell to one of two training images** — creating a checkerboard-like mixing pattern that distributes information from both images evenly across the spatial dimensions. **How Does GridMix Work?** - **Grid**: Divide the image into an $n imes n$ grid of cells. - **Assignment**: Randomly assign each cell to image $A$ or image $B$ with probability $lambda$. - **Mix**: Fill each cell with the corresponding region from the assigned image. - **Labels**: Mixed proportionally to the number of cells assigned to each image. **Why It Matters** - **Spatial Distribution**: Unlike CutMix (single contiguous region), GridMix distributes both images across the entire spatial extent. - **Multiple Regions**: Forces the model to handle multiple disjoint regions from each class simultaneously. - **Complementary**: Can be combined with other augmentation strategies. **GridMix** is **checkerboard image mixing** — distributing both images across a grid for spatially diverse data augmentation.