# U-Net analytical reference

Run `python examples/unet.py` or `python examples/unet.py --json`.
No install is needed. `requirements-unet.txt` intentionally lists no packages.

The deterministic checks follow the original **valid convolution** network:
572 input → 388 prediction, four center-cropped concatenation skips, and the
boundary-weight formula with supplied distances and class weight. Regions:
`geometry`, `crop`, `boundary`.

Pinned source: [Ronneberger, Fischer, Brox, arXiv:1505.04597v1](https://arxiv.org/pdf/1505.04597v1),
Figure 1 and Equations 1–2. The source-crop manifest records the PDF SHA-256.

This is arithmetic for teaching, not a convolution implementation, trained U-Net,
augmentation pipeline, biomedical validation or timing benchmark. Distances must
be computed from instance boundaries by a real training pipeline. Modern padded
U-Nets, BatchNorm variants and multiclass heads differ from this reference.
