API: models

Network architectures. Module: autoStructN2V.models.

FlexibleUNet (models/unet.py)

The configurable U-Net both branches train. Architecture options map 1:1 to recipe keys:

Recipe key	Effect
features, num_layers	Width and depth
use_resize_conv, upsampling_mode	Resize-convolution upsampling (ResizeConvolution) instead of transposed conv — prevents checkerboard artifacts
remove_top_skip	N2V2: drop the top skip connection
use_blurpool	N2V2: anti-aliased downsampling (MaxBlurPool2d)
activation	'relu' (publication) / 'elu' (package default)
norm_type, num_groups	'batch' or 'group' normalization
init_scale	Weight-init scaling

Building blocks exported alongside: MaxBlurPool2d, ResizeConvolution.

Convenience constructors: create_standard_unet(...), create_resize_conv_unet(...).

AutoStructN2VModel (models/auto_struct_n2v.py)

Wrapper that builds a branch model from a recipe: .create_stage1_model(...) (N2V branch) and .create_stage2_model(...) (StructN2V branch) — the stage names are legacy, the branches are current.

Factory (models/factory.py)

• create_model(stage, **arch_kwargs) — build a model for 'n2v'/'structn2v' (legacy 'stage1'/'stage2' accepted)
• create_model_from_config(config, stage) — build directly from a validated pipeline config

Example

from autoStructN2V.models.factory import create_model_from_config
from autoStructN2V.pipeline import validate_config

config = validate_config({"input_data": "stack.tif"})
model = create_model_from_config(config, stage="structn2v")

See Architecture & Training Recipe for what the publication recipe sets and why.