Quadtree Tensor Compression for Sparse Neural Network Data

The tensor is divided into multiple blocks, and each block is further segmented into sub-blocks. A quadtree data structure is constructed for each block to represent the distribution of non-zero cells at different levels of granularity. This segmentation allows the compression method to adapt to local characteristics of different regions within the tensor, improving compression efficiency by focusing computational resources on blocks with higher non-zero cell density while using more compact representations for blocks with predominantly zero values.

The compression method dynamically selects different compression modes (first mode, second mode, or third mode) based on the actual distribution characteristics of non-zero cells in each block. The mode selection is determined by analyzing the quadtree structure and comparing metrics such as the ratio of non-zero cells to total cells. This dynamic adaptation allows the system to optimize data transfer speed for each block individually, using more aggressive compression for zero-dominated regions while maintaining higher fidelity for regions with significant non-zero content.

Different quantization strategies are applied to different blocks based on their local characteristics. Blocks with high non-zero cell density undergo less aggressive quantization or use higher precision representations, while blocks with predominantly zero values use more aggressive quantization schemes. The quadtree structure enables this local quality differentiation by providing a hierarchical representation that captures the spatial distribution of non-zero cells, allowing the system to apply appropriate quantization levels to each region to minimize overall data loss while reducing total data size.