cellmap_flow.dashboard.routes.model_advice

Advice about a served model’s normalization and postprocessing.

Nothing in cellmap-flow records what a model expects: metadata.json carries geometry and provenance only, and the input/output handling is set by convention plus whatever happens to be toggled in the UI. A wrong choice degrades the result silently rather than erroring, which is a bad failure mode.

This endpoint combines two sources into a verdict the UI can show:

• the running server’s model info (/__control__/model_info), which says what activation the model’s output already has: read off its network’s graph (the op that ends it), or else observed by its output probe on noise, a guess, and

• the model’s declared metadata (channel count, name, training framework).

It never changes anything – callers decide whether to apply the suggestion. Nor does it undo what the Finetune tab’s ranking set: the input normalization the ranking fitted to the paint for a model (Session.rank_fits) counts as matching, not as one to suggest the declared one back over, and an affinity model’s chain into a segmentation is offered cut at the threshold it fitted.

Attributes

logger
model_advice_bp

Functions

model_metadata(model_config) → dict	Best-effort metadata for a model config, from whatever source has it.
raw_dtype_and_range(dataset_path)	The dtype and value range of the raw data, for scaling it to the model.
input_norm_suggestion(model_config, meta, raw_dtype, ...)	What model_config's input normalization should be on raw data of
model_advice()	Per-model verdicts on the currently configured input/output handling.

Module Contents

cellmap_flow.dashboard.routes.model_advice.logger

cellmap_flow.dashboard.routes.model_advice.model_advice_bp

cellmap_flow.dashboard.routes.model_advice.model_metadata(model_config) → dict

Best-effort metadata for a model config, from whatever source has it.

cellmap_flow.dashboard.routes.model_advice.raw_dtype_and_range(dataset_path)

The dtype and value range of the raw data, for scaling it to the model.

Integer data has an exact range from its dtype. Float data does not – the dtype permits anything – so sample percentiles instead, the same way the raw layer’s auto-contrast does, on the channel the chain picks. Returns (dtype_name, (lo, hi)); either part may be None when it could not be determined.

cellmap_flow.dashboard.routes.model_advice.input_norm_suggestion(model_config, meta, raw_dtype, raw_range)

What model_config’s input normalization should be on raw data of raw_dtype spanning raw_range: suggest_input_norm’s suggestion, and the declaration it holds the chain to (_declared_input_norm), or None when the model declares none. meta is model_metadata’s for it ({} for a job whose model is not among the session’s).

The page shows the suggestion, and a model’s start sets it on a chain with no input normalization when it is the model’s declaration or the convention for its kind of model (routes.pipeline.set_declared_input_norm).

cellmap_flow.dashboard.routes.model_advice.model_advice()

Per-model verdicts on the currently configured input/output handling.

{"success": true, "models": [...], "starting", "digest"}: an entry per served model; starting, whether a model’s start is still on its way (services/launch.py launching: its job may be listed already while its start is still to set the chain’s postprocessing, routes.pipeline.set_needed_postprocess), and digest, the chain’s (PipelineSpec.digest). The page asks again while starting, and reads the chain again when the digest moves: a model’s start sets the postprocessing its output needs once its server answers, after the Submit that started it.