cellmap_flow.process_chain

This process’s normalization + postprocessing chain: process_chain().

One chain per process. In the dashboard it is the chain the user last submitted, which the finetune manifest and the exported YAML read as spec. In an inference server or a blockwise worker it is the fallback for a layer or caller that brings no chain of its own: always empty for a server started from the CLI, and set by the user scripts that call process_chunk themselves (ImageDataInterface’s docstring says they rely on it).

• input_norms and postprocess are the live op instances, which can hold state (SimpleBlockwiseMerger’s equivalences), so they are replaced only by set() and never rebuilt from the spec behind a caller’s back.

• input_norm_config and postprocess_config are the steps as the dashboard received them; spec reads them.

• set(spec, built=None) is the one way to change the chain.

Classes

ProcessChain

The chain; see the module docstring.

Functions

process_chain(→ ProcessChain)

This process's chain.

Module Contents

class cellmap_flow.process_chain.ProcessChain

The chain; see the module docstring.

input_norms = []

postprocess = []

input_norm_config

postprocess_config

property spec: cellmap_flow.pipeline_spec.PipelineSpec

The chain currently configured, as data. Derived on every read.

Per chain: the *_config steps when set, otherwise the live instances’ to_dict(). Never rebuilds the live instances, which can hold state (SimpleBlockwiseMerger’s equivalences).

set(spec: cellmap_flow.pipeline_spec.PipelineSpec, built=None) → None

Replace the configured chain: both live chains and both configs.

built is the (input_norms, postprocess) instances for spec when the caller already has them; otherwise they are built here. Everything is built before anything is assigned, so a chain that fails to build leaves the previous one in place.

output_dtype(model_output_dtype, postprocess=None)

The dtype a chain hands to the client.

postprocess=None means this chain’s; the inference server passes the chain of the layer being served. The last step that declares a dtype decides, since the steps run in order: taking the first picked e.g. SigmoidPostprocessor’s float32 ahead of a trailing AffinityPostprocessor’s uint64, which both advertised the wrong dtype in the zarr metadata (neuroglancer: “Data type not compatible with segmentation layer”) and cast uint64 label ids through float32, corrupting any id above 2**24.

cellmap_flow.process_chain.process_chain() → ProcessChain

This process’s chain.

Read it at call time, never into a module global: tests swap in a fresh chain per test, and a copy taken at import would outlive them.