API: datasets

Patch sampling and blind-spot masking. Module: autoStructN2V.datasets.

Class Overview

Class	Role
BaseNoiseDataset	Shared patch-extraction and normalization machinery
TrainingDataset	Random patches + blind-spot masking + (N2V branch) augmentation
ValidationDataset	Held-out patches with the same masking, fixed sampling
TestDataset	Full-slice access for evaluation

In normal use you don't construct these directly — create_routed_dataloaders (see pipeline API) builds them from the validated config and the routed branch's kernel.

The Masking Step

TrainingDataset.apply_mask(patch, mask, mask_strat, prediction_kernel=None) corrupts a patch for blind-spot training:

1. mask marks the positions to blank (prediction centers stamped with the branch kernel, built by create_full_mask)
2. mask_strat selects the replacement values — the publication setting is UPS (strategy 3): a uniform sample from the 5×5 neighborhood excluding masked positions
3. The loss is later computed only at the prediction centers

Notes

• mask_percentage counts prediction centers, not total corrupted pixels — a StructN2V kernel blanks several pixels per center
• Augmentation (flips + 90° rotations) is applied only when the recipe enables it; the StructN2V branch keeps it off to preserve mask–noise alignment
• The dataset classes still carry legacy ROI-selection parameters from the retired design; the pipeline hardwires use_roi=False and the config translator drops the old keys

Related

• Architecture & Training Recipe — how masking fits the training
• Masking API — kernel construction