cellmap_flow.finetune.finetuned_model_templates

Templates for generating finetuned model YAML configurations.

This module provides functions to auto-generate the YAML config for serving finetuned models using FinetuneModelConfig (type: finetune).

Attributes

logger

FINETUNED_MODEL_YAML_MARKER

PLACEHOLDER_DATA_PATHS

Functions

generate_finetuned_model_yaml(→ pathlib.Path)

Generate .yaml configuration for serving a finetuned model.

generate_current_config_yaml(→ str)

Build YAML text snapshotting the dashboard's current live server config

Module Contents

cellmap_flow.finetune.finetuned_model_templates.logger
cellmap_flow.finetune.finetuned_model_templates.FINETUNED_MODEL_YAML_MARKER = 'FINETUNED_MODEL_YAML:'
cellmap_flow.finetune.finetuned_model_templates.PLACEHOLDER_DATA_PATHS
cellmap_flow.finetune.finetuned_model_templates.generate_finetuned_model_yaml(lora_adapter_path: str = None, base_model_dict: dict = None, model_name: str = None, output_path: pathlib.Path = None, data_path: str = None, queue: str = None, charge_group: str = None, json_data: dict = None, scale: str = None, weights_path: str = None, resample: bool = False, input_voxel_size=None, output_voxel_size=None) → pathlib.Path

Generate .yaml configuration for serving a finetuned model.

The generated YAML uses type: finetune, which delegates to FinetuneModelConfig. This loads the base model via its own ModelConfig, applies the LoRA adapter, and serves the result.

Parameters:

  • lora_adapter_path – Path to the saved LoRA adapter directory

  • weights_path – Instead of an adapter, a full state dict from a –lora-r 0 (full finetune) run. Exactly one of the two.

  • base_model_dict – Dict describing the base model (from model_config.to_dict())

  • model_name – Name of the finetuned model

  • output_path – Where to write the .yaml file

  • data_path – Path to actual dataset (REQUIRED - no placeholders)

  • queue – LSF queue name; None is the site’s default queue

  • charge_group – LSF charge group; None is the site’s default one

  • json_data – Optional dict with input_norm and postprocess from base model

  • scale – The level of a multiscale data_path to serve (“s1”), or None for none. With none, the server picks the level by the model’s input voxel size, which is how the trainer read the same data_path. This was always “s0”, so a model trained on s1 of a multiscale group was served s0.

  • resample – the model was trained on the raw resampled to its voxel size (its volume’s resample), so serve it that way too

  • input_voxel_size – the voxel sizes the finetune was trained at, when not the base model’s (run_outputs); the served model declares them

  • output_voxel_size – the voxel sizes the finetune was trained at, when not the base model’s (run_outputs); the served model declares them

Returns:

Path to the generated YAML file

cellmap_flow.finetune.finetuned_model_templates.generate_current_config_yaml(models: list, data_path: str, queue: str = None, charge_group: str = None, walltime: str = None, json_data: dict = None, resample: bool = False) → str

Build YAML text snapshotting the dashboard’s current live server config (whatever models/normalization/postprocessing/queue are active right now), so it can be handed back to cellmap_flow yaml later.

Unlike generate_finetuned_model_yaml (which describes one specific finetune job), this is a general “export what’s currently running” dump – it doesn’t write to disk itself, callers decide where the result goes (a download response, a file, etc).

Parameters:

  • models – list of model entry dicts (each from ModelConfig.to_dict())

  • data_path – current dataset path

  • queue – LSF queue name; None is the site’s default queue

  • charge_group – LSF charge group; None is the site’s default one

  • walltime – LSF run limit (“HH:MM” or minutes); omitted when unset, so the reloaded config falls back to the built-in default

  • json_data – dict with “input_norm”/”postprocess” keys reflecting the currently active normalization/postprocessing, if any

  • resample – the dashboard’s Resample setting, written either way

Returns:

YAML text (str) ready to write to a file.