cellmap_flow.models.configs.cellmap

CellMapModelConfig: a model that cellmap_models exported, from its folder.

The folder’s metadata gives the shapes, voxel sizes and channel names; the model is its TorchScript export.

Attributes

logger

Classes

CellMapModelConfig

Configuration class for a CellmapModel.

Functions

`runs_at`(model, input_shape, output_shape) → bool	Whether the TorchScript `model` runs on an input of `input_shape`
`export_finetune_modes`(file_names) → tuple	How a cellmap_models export whose top-level files are `file_names`
`metadata_geometry`(metadata)	`ModelConfig.declared_geometry` from an export's metadata (`metadata`

Module Contents

cellmap_flow.models.configs.cellmap.logger

cellmap_flow.models.configs.cellmap.runs_at(model, input_shape, output_shape) → bool

Whether the TorchScript model runs on an input of input_shape voxels and answers output_shape.

cellmap_models traces the export (torch.jit.trace) at whichever shape the exporter passed, and a traced UNet’s crops are baked in, so it runs only at that shape: fed another, it either fails (jrc_mus-livers_16nm_to_8nm_mito, traced at its inference 288³, fails at its training 216³ with “padded input size (1 x 1 x 1)”) or answers the baked-in output (mito_aff_unet_setup_16, traced at its training 178³ → 56³, answers 56³ to a 378³ input). The metadata records both shapes and not which one the export was traced at, so this asks the model.

cellmap_flow.models.configs.cellmap.export_finetune_modes(file_names) → tuple

How a cellmap_models export whose top-level files are file_names can be finetuned, from their names alone: LoRA and full with a model.pt2, which cellmap_models’ train() unflattens into a module adapters attach to; full only without one, as train() then hands the trainer the TorchScript, a compiled network that takes no adapter.

The one rule for an export’s folder (CellMapModelConfig.finetune_modes), its Hugging Face repository (HuggingFaceModelConfig’s) and the model lists’ facts (hf_catalog.weight_facts). fly_organelles’ copies under /nrs/cellmap/models/saalfeldlab (model.pt and model.ts; their Hugging Face copies have a model.pt2) and cellmap/jrc_mus-livers_16nm_to_8nm_mito (model.pt, model.ts and model.onnx) have none: offered LoRA, their jobs failed at their start.

cellmap_flow.models.configs.cellmap.metadata_geometry(metadata): ModelConfig.declared_geometry from an export’s metadata (metadata has its fields as attributes): its training output, when the metadata gives no other shape (_other_shapes), so that it is the one the export runs at; else None, as when a field is missing. setup_16’s metadata gives inference shapes too, and its export runs at the training ones; jrc_mus-livers_16nm_to_8nm_mito’s runs at its inference ones.

class cellmap_flow.models.configs.cellmap.CellMapModelConfig(folder_path, name=None, scale=None)

Configuration class for a CellmapModel.

cli_name = 'cellmap'

finetunable = True

cellmap_model

name = None

scale = None

finetune_modes(): export_finetune_modes of its folder’s files, read from their names without loading them: the model list asks on every poll. Both while the folder cannot be listed.

declared_geometry(): Its export’s metadata.json’s (metadata_geometry), read without the model.

declared_input_norm(data_range=None): Its export’s metadata.json’s input_normalization (metadata_input_norm). cellmap_models’ parsed metadata leaves that field out, so the file is read for it, once.

to_dict(): This config as a model entry, which registry.build_model rebuilds.