| # HeatTok Semantic Patch Cache |
|
|
| Precomputed `.pt` caches for HeatTok. Use with `HEATTOK_SEMANTIC_CACHE_DIR=/path/to/cache`. |
|
|
| Filename pattern: `{image_hash}_g1_s28.pt` or `{image_hash}_g1_gd1_s28.pt` |
|
|
| --- |
|
|
| ## semantic_patch_cache_vrsbench |
| |
| - **Dataset:** VRSBench (512×512) |
| - Caches do **not** store precomputed global tokens or patch orientations. |
| - **Gaussian parameters** and patch metadata **are** stored. |
| - **No need to regenerate** `.pt` files — HeatTok computes global tokens and orientations online at load time. |
| |
| --- |
| |
| ## semantic_patch_cache_earthvqa |
|
|
| - **Dataset:** EarthVQA (1024×1024) |
| - Caches use **global downsample divisor 1** (`gd1`). |
| - To use **divisor 8** (`gd8`), no regeneration needed — global branch is rebuilt online from the image processor. |
|
|
| --- |
|
|
| ## Usage |
|
|
| ```bash |
| export HEATTOK_SEMANTIC_CACHE_DIR=/path/to/semantic_patch_cache_vrsbench # or earthvqa |
| ``` |
|
|
| ```yaml |
| use_semantic_patches: true |
| global_downsample: true |
| semantic_patch_size: 28 |
| ``` |
|
|
|
|
|
|
