API: pipeline

The routed runner, configuration handling, and data splitting. Module: autoStructN2V.pipeline.

Runner (pipeline/runner.py)

run_pipeline(config) -> dict

The package's main entry point: validate config → resolve the route → train one model → denoise the stack. Returns a summary dict with branch, route_reason, route_metrics, mask_path, model_path, denoised_stack, experiment_dir, final_results_dir — or halted, halt_reason, mask_review_dir when extractor_input='compare' stopped the run for review.

resolve_route(config, stack, dirs=None, verbose=False) -> RouteDecision

Resolves the mask source into a routing decision without training: runs the extractor (source='extractor'), builds a center kernel ('center'), or loads a file kernel ('file').

load_mask_from_file(path, verbose=False) -> np.ndarray

Loads and validates a .npy boolean kernel.

Configuration (pipeline/config.py)

validate_config(config) -> dict

Fills every unspecified key with a documented default, validates types/ranges/paths, and auto-translates legacy two-stage configs (stage1 → recipes.n2v, stage2 → recipes.structn2v, mask block derived). The retired ring-Otsu mask_source='stage1' path raises; retired ROI-selection keys are dropped.

create_output_directories(config) -> dict

Creates the experiment directory tree and returns its paths.

Defaults dictionaries

TOP_DEFAULTS, N2V_RECIPE_DEFAULTS, STRUCTN2V_RECIPE_DEFAULTS, EXTRACTOR_DEFAULTS — the source of truth for every default in the Configuration guide.

Data (pipeline/data.py)

split_stack_indices(num_slices, split_ratio=(0.7, 0.15, 0.15), seed=None, verbose=False)

Deterministic train/val/test split of z-slice indices.

create_routed_dataloaders(config, recipe, single_kernel, stack, slice_indices, verbose=False)

Builds the train/val/test dataloaders for the routed branch: patch sampling, blind-spot masking with the branch kernel, normalization, and (N2V branch only) augmentation.

Example

from autoStructN2V.pipeline import run_pipeline, validate_config, resolve_route

config = validate_config({
    "input_data": "stack.tif",
    "mask": {"source": "extractor", "extractor": {"bg_side": "light"}},
})

# peek at the route without training:
import tifffile, numpy as np
stack = tifffile.imread("stack.tif").astype(np.float32)
decision = resolve_route(config, stack, verbose=True)
print(decision.branch, decision.reason, decision.message)

summary = run_pipeline(config)