Architecture Overview¶

DNANet is organized into six packages, each with a clear responsibility:

dnanet/
├── core/          Domain models (alleles, markers, panels, annotations)
├── data/          Data loading, parsing, preprocessing, strategies
├── models/        Neural network architectures (stateless nn.Module)
├── modules/       PyTorch Lightning modules (training logic)
├── evaluation/    Metrics, allele calling, visualization
└── tasks/         CLI task runners (train, evaluate, cross-validate)

Package Dependencies¶

The dependency graph flows strictly downward — no circular imports:

tasks/
  ↓
modules/     evaluation/
  ↓              ↓
models/      data/
  ↓           ↓
core/       core/

• core depends on nothing (pure domain models + stdlib)
• data depends on core + NumPy/SciPy
• models depends on core + PyTorch
• modules depends on models + Lightning
• evaluation depends on core + data + NumPy
• tasks depends on everything (orchestration layer)

Key Design Decisions¶

1. Separation of Architecture from Training¶

Neural network architectures (models/) are pure nn.Module classes with no training logic. Training loops, optimizers, metrics, and logging live in Lightning modules (modules/). This allows:

• Reusing the same UNet architecture for segmentation or reconstruction
• Unit-testing architectures without Lightning overhead
• Swapping training strategies without touching model code

2. Strategy Pattern for Kit & Dataset Variance¶

Forensic DNA kits (PPF6C, GlobalFiler) and datasets (NFI R&D, ProvedIt) have different conventions for size standards, file naming, and annotations. Rather than scattering if kit == "PPF6C" conditions, each variant is a Strategy class. Kit scaling strategies and dataset strategies are injected directly into the data pipeline.

3. Hydra Composition over Monolithic Config¶

Configuration is split into independent groups (data, model, training, evaluation, logging) that are composed at runtime. This avoids the combinatorial explosion of maintaining separate config files for every experiment combination.

4. Lazy Loading for HIDImage¶

HID files are parsed on first access to the .data property, not at construction time. This means scanning a directory of 800 files is instant; the I/O cost is paid only when data is actually needed.

Module Map¶