uiol;' (#154)

[DeepKnowledge1]

Pull Request: Production-ready TensorRT/INT8 export, ultra-light PatchCore, and Production Autopilot

Summary

This pull request merges feat/tensorrt-int8-light-patchcore into main. It adds production-oriented anomaly-detection capabilities, improves deployment workflows, strengthens benchmarking, and makes the project easier for new users to evaluate.

What changed

Native TensorRT and INT8 export

• Added native TensorRT engine export with FP16 and FP32 support.
• Added calibrated INT8 TensorRT export using real normal calibration images.
• Added dynamic batch profiles and configurable TensorRT workspace limits.
• Added calibration-cache support and engine validation.
• Added automatic conversion support for compact PaDiM and PatchCore deployment artifacts.
• Updated the configuration template and deployment documentation with the new export parameters.

Ultra-light PatchCore

• Added a production-oriented PatchCore implementation designed for bounded memory and predictable inference.
• Added deterministic k-center coreset selection.
• Added configurable memory-bank limits, pooled patch grids, and chunked nearest-neighbor search.
• Added support for CLI training, validation, testing, inference, export, and configuration workflows consistent with PaDiM.
• Restored PatchCore localization output and corrected threshold and boundary-color behavior.

Production Autopilot

• Added a hardware-aware CLI workflow that compares PaDiM and ultra-light PatchCore.
• Added algorithm-specific threshold calibration using the complete labeled test split by default.
• Added image-level and pixel-level AUROC reporting when valid targets are available.
• Added localization-health diagnostics, including anomaly coverage and normal-image false-positive localization.
• Added median and P95 latency profiling.
• Added model selection based on accuracy, latency, and production constraints.
• Added a self-contained HTML deployment dashboard, deployment manifest, localization report, and selected model package.

Example:

anomavision autopilot \
  --config config.yml \
  --padim_model ./distributions/padim/bottle/anomav_exp/model.pt \
  --patchcore_model ./distributions/patchcore/bottle/anomav_exp/model.pt \
  --device cpu \
  --validation_split 1.0 \
  --target_latency_ms 50 \
  --output_dir ./production_package

Documentation and adoption improvements

• Added a five-minute CPU quickstart that does not require CUDA.
• Added copy-ready examples for CPU PaDiM, ultra-light PatchCore, and TensorRT INT8 export.
• Improved README navigation, visual explanations, Autopilot guidance, and CI visibility.
• Added reproducible benchmark instructions and P95 latency reporting.
• Updated the public repository description and GitHub topic tags for improved discoverability.
• Added direct links to the live demo, quickstart, examples, and benchmark methodology.

Release preparation

• Bumped the package version from 3.4.0 to 3.5.0.
• Updated uv.lock to match the new package version.
• The existing release workflow is configured to build the package, publish it to PyPI through trusted publishing, and attach artifacts to the GitHub release.

Validation

The complete repository validation suite passes:

black       Passed
isort       Passed
flake8      Passed
pytest      Passed

The final branch is clean and synchronized with the remote feature branch.

Compatibility and deployment notes

• CPU installation and inference remain supported.
• CUDA and TensorRT export require the corresponding NVIDIA software stack.
• INT8 export requires real, representative normal calibration images.
• TensorRT engines should be validated on the target GPU and TensorRT runtime.
• PaDiM and PatchCore use different score scales; their thresholds must be calibrated separately.
• Existing users should review the updated configuration template before upgrading.

Recommended merge and release sequence

1. Review and merge this pull request into main.
2. Confirm the package version is 3.5.0 on main.
3. Create and publish a GitHub release tagged v3.5.0.
4. Allow the release workflow to build and publish the package to PyPI and attach the artifacts.

Checklist

• Native TensorRT FP16/FP32 export added.
• Calibrated TensorRT INT8 export added.
• Ultra-light PatchCore added and documented.
• Production Autopilot added.
• Threshold and localization behavior corrected.
• README and deployment documentation updated.
• CPU quickstart and runnable examples added.
• P95 benchmark reporting added.
• Package version updated to 3.5.0.
• Full pre-commit validation passed.
• Merge into main.
• Create GitHub release v3.5.0.