Upsampling in AnySat with "dense" argument instead of "patch" argument

[danielkiv]

According to the author of AnySat, currently the project uses the "patch" argument (line 403 of onthefly_anysat.py) but the best way to use the model is through the "dense" argument as all segmentation and other dense tasks evaluations of the model are performed under this setting according to AnySat paper. This will lead to a much bigger embedding matrix (one embedding per pixel) but it is possible down-sample if needed.

[Dinghye]

Summary of changes

This update adjusts AnySat output handling in a compatibility-preserving way, while keeping the changes localized to the AnySat adapter rather than changing the framework-level OutputSpec design.

1. grid now defaults to dense instead of patch

• OutputSpec.grid() now uses AnySat’s native dense output by default
• This is closer to the way dense tasks are handled in the AnySat paper / official implementation
• In rs-embed, it is still exposed through the standard (D, H, W) grid output

2. The previous pooled default behavior is preserved

• OutputSpec.pooled() still defaults to mean/max pooling over the AnySat patch grid
• This keeps historical behavior and pooled vector dimensionality stable, avoiding an immediate breaking change

3. Added AnySat-specific configuration options

• grid_feature_mode="dense" | "patch"
• pooled_source="patch" | "tile"

These options are only interpreted by the AnySat adapter and do not affect other models.

Current behavior

• OutputSpec.grid() now maps to AnySat output="dense" by default
• OutputSpec.grid(..., grid_feature_mode="patch") restores the previous patch-grid behavior
• OutputSpec.pooled() still defaults to patch-grid pooling
• OutputSpec.pooled(..., pooled_source="tile") switches to AnySat’s native tile embedding

API handling

The public API still follows the existing direct-keyword style. Users do not need to manually construct model_config:

get_embedding("anysat", ..., output=OutputSpec.grid(), grid_feature_mode="patch")
get_embedding("anysat", ..., output=OutputSpec.pooled(), pooled_source="tile")

The export path supports the same settings through:

ExportModelRequest.configure("anysat", grid_feature_mode="patch")
ExportModelRequest.configure("anysat", pooled_source="tile")

Why this design

• grid -> dense: aligns better with the official AnySat dense-task usage
• pooled does not default to tile: avoids changing pooled dimensionality / semantics by default
• all changes are contained inside the AnySat adapter, without introducing a new framework-wide spec or changing OutputSpec

Documentation updates

The AnySat docs were updated to clarify that:

• grid now defaults to dense
• pooled still defaults to patch pooling, but tile is now supported
• corresponding environment variables were added/documented:
  • RS_EMBED_ANYSAT_GRID_MODE
  • RS_EMBED_ANYSAT_POOLED_SOURCE

Tests

Added focused AnySat tests covering four paths:

• grid defaults to dense
• grid_feature_mode="patch" falls back to patch mode
• pooled keeps patch pooling by default
• pooled_source="tile" uses native tile output

Current focused test result: 4 passed

Files touched

• onthefly_anysat.py
• anysat.md
• models_reference.md
• test_anysat_output_modes.py