cellmap_flow.models.geometry_cache

Get a model’s geometry without building the model, when possible.

Sizing an annotation volume needs five numbers: read_shape, write_shape, the two voxel sizes, and the output channel count. Reading them from ModelConfig.config costs a full model load – for a script model that executes the config file, which downloads weights, loads adapters, and runs a validation forward pass. One measured session spent 43s of a 47s “create annotation volume” there, on the dashboard’s CPU, for numbers the inference server already had.

Three sources, cheapest first:

a running inference server, which already holds the model;
this cache, keyed so that it invalidates itself when the model changes;
building the model, which then populates the cache. Not for a model that runs in its own environment (models.envs): its packages are there, not here (build_here).

Reading the declarations statically is deliberately not among them. The script contract allows geometry to depend on the model – a DaCapo script can say output_voxel_size = Coordinate(model.scale(voxel_size)) – so parsing literals would silently produce wrong numbers for those scripts rather than failing.

Attributes

`logger`
`CACHE_PATH`
`MAX_ENTRIES`

Functions

`cache_key`(model_config)	A key that changes when the model does, or None if we cannot make one.
`load_cached_geometry`(model_config)	Geometry remembered for this exact model, as a geometry_stand_in, or None.
`store_geometry`(model_config, config)	Remember the geometry of a model we just paid to build.
`build_here`(model_config)	`model_config.config`, built in this process.
`resolve_model_geometry`(model_name, model_config)	The five geometry fields, by the cheapest route that can supply them.

Module Contents

cellmap_flow.models.geometry_cache.logger

cellmap_flow.models.geometry_cache.CACHE_PATH

cellmap_flow.models.geometry_cache.MAX_ENTRIES = 200

cellmap_flow.models.geometry_cache.cache_key(model_config)

A key that changes when the model does, or None if we cannot make one.

Only keys that invalidate themselves are allowed. A script is keyed by path and mtime, so editing it misses the cache. A HuggingFace repo is keyed by an explicit revision, which is immutable; without one the revision is whatever main points at today, so those are not cached rather than risk serving geometry for a model that has since moved.

cellmap_flow.models.geometry_cache.load_cached_geometry(model_config): Geometry remembered for this exact model, as a geometry_stand_in, or None.

cellmap_flow.models.geometry_cache.store_geometry(model_config, config)

Remember the geometry of a model we just paid to build.

config is its ModelConfig.config, or anything with the same geometry attributes, a ModelGeometry included.

cellmap_flow.models.geometry_cache.build_here(model_config)

model_config.config, built in this process.

Refused for a model that runs in another environment (its env, or its type’s default_env): this process lacks its packages, or has other versions of them (the default environment’s cellpose 3 for a Cellpose-SAM model), so building it here fails or, worse, half-works. Its geometry comes from its running server.

Raises:: ModelEnvError – the model runs in its own environment.

cellmap_flow.models.geometry_cache.resolve_model_geometry(model_name, model_config)

The five geometry fields, by the cheapest route that can supply them.

Returns an object exposing read_shape, write_shape, input_voxel_size, output_voxel_size and output_channels – either a stand-in or, on the build path, the real ModelConfig.config. Raises ModelEnvError for a model that runs in its own environment when neither its server nor the cache knows its geometry.