XCalib Code Index¶
Open this when locating a public API, model, runtime backend, data flow, continual-learning component, GUI surface, or validation command.
- Package map
- Inference and calibration flow
- Training and continual flow
- Distribution and documentation
Package map¶
| Path | Responsibility |
|---|---|
src/xcalib/__init__.py |
Public imports and package version. |
src/xcalib/engine/matcher.py |
Main matcher construction and inference facade. |
src/xcalib/engine/exporter.py |
ONNX export and parity support. |
src/xcalib/engine/trainer.py |
Shared training behavior. |
src/xcalib/engine/trt.py |
TensorRT runtime integration. |
src/xcalib/models/registry.py |
Canonical model registration and construction. |
src/xcalib/cfg/ |
Packaged model/site configuration. |
src/xcalib/protocol.py |
Public input and record protocol. |
src/xcalib/calibration_refine.py |
Calibration refinement and candidate selection. |
src/xcalib/oneshot/ |
One-shot adaptation sessions, memory, and pseudo-labels. |
src/xcalib/continual/ |
Continual jobs, revisions, training, stores, and provenance. |
src/xcalib/data/ |
Crop and HDF5 cache loading. |
src/xcalib/hub/ |
Public read-side weights and dataset retrieval. |
src/xcalib/gui/ |
Review application and scene model. |
tests/smoke/ |
Fast import and forward-pass checks. |
tests/unit/ |
Component and contract tests. |
tests/integration/ |
Multi-component build, training, and one-shot checks. |
Inference and calibration flow¶
The CLI and public imports construct a matcher through the engine and model registry. Packaged configuration selects the model/site behavior. Assignment, calibration refinement, visualization, and optional runtime wrappers consume the protocol-defined inputs without introducing backend-specific public APIs.
Training and continual flow¶
HDF5 loaders feed shared training code and paper scripts. One-shot adaptation uses session-local memory and pseudo-labels. Continual modules record revisions, jobs, data provenance, training state, and publication-ready records while preserving atomic state transitions.
Distribution and documentation¶
pyproject.toml owns package metadata, dependencies, Pixi tasks, build rules,
and the dynamic version path. mkdocs.yml owns the public documentation portal.
scripts/check_public_docs.py guards tracked files against personal paths,
credentials and Hub write-side operations.