reflecting the EfficientAD integration, performance work, calibration, localization, Autopilot support, and testing fixes - #166

🔗 Related Issue

Fixes #

📝 Description

This PR adds EfficientAD support to AnomaVision and integrates it into the existing anomaly-detection pipeline alongside PaDiM and PatchCore.

🚀 EfficientAD

• Added EfficientAD as a supported anomaly-detection algorithm.

• Integrated EfficientAD with the existing algorithm abstraction used by PaDiM/PatchCore.

• Reused the existing AnomaVision data loading and preprocessing pipeline.

• Added batched training/inference to avoid unnecessary repeated computation.

• Optimized teacher feature extraction and removed unnecessary inference-time computation.

• Added calibrated anomaly thresholds based on normal training images only.

• Added anomaly-map generation for pixel-level localization.

• Reused the existing post-processing and visualization pipeline for:

  • anomaly heatmaps
  • localization masks
  • contours
  • bounding boxes
  • overlay visualizations

🔌 Deployment & Export

• Added EfficientAD support to the existing model/export workflow.
• Maintained compatibility with the existing ONNX-based detection flow.
• Preserved the existing PaDiM/PatchCore inference architecture rather than introducing a separate EfficientAD pipeline.
• Added EfficientAD support to Production Autopilot.
• Autopilot can compare PaDiM, PatchCore, and EfficientAD candidates.
• EfficientAD calibration data is preserved with the model artifact.

📊 Production Autopilot

• Restored and preserved the rich production_autopilot_report.html dashboard.

• Added EfficientAD to candidate comparison and model selection.

• Report includes:

  • Image AUROC
  • Pixel AUROC
  • median latency
  • P95 latency
  • anomaly coverage
  • normal false-positive localization rate
  • threshold
  • selected model
  • deployment artifact
  • localization health
  • reproducibility information

• Restored the existing timing/performance summary in the detection pipeline.

🐛 Fixes

• Fixed EfficientAD training failure caused by PyTorch inference-mode tensors being used in autograd.
• Fixed JSON serialization of NumPy/PyTorch scalar metrics in the Autopilot deployment manifest.
• Restored Autopilot report/test compatibility.
• Preserved existing helper APIs such as _format_metric.
• Fixed report metric naming/capitalization expected by the test suite.

🔄 Type of Change

• 🐛 Bug fix (non-breaking change which fixes an issue)
• 🚀 New feature (non-breaking change which adds functionality)
• 💥 Breaking change (fix or feature that would cause existing functionality to not work as expected)
• 📖 Documentation update
• 🏗️ Infrastructure / CI/CD update

🧪 Hardware & Matrix Testing

I have successfully built and tested this code using uv on:

• anomavision[cpu] (Standard/Edge)
• anomavision[cu121] (CUDA 12.1)
• anomavision[cu124] (CUDA 12.4)
• anomavision[cu118] (CUDA 11.8)

Host OS used for testing:

• Linux / Ubuntu
• Windows (Native or WSL2)
• macOS

🧪 Validation

• Existing Autopilot unit tests updated/fixed.
• EfficientAD-specific integration changes covered.
• Autopilot report generation validated.
• Deployment manifest generation validated.
• Full uv run pytest suite verified on all supported environments.
• Bottle dataset benchmark completed for PaDiM vs PatchCore vs EfficientAD.
• Full ONNX/OpenVINO/TensorRT/Hailo hardware validation completed.

✅ Developer Checklist

• My code follows the core style guidelines of this project (Ruff/Black formatting).
• I have run uv run pytest and all unit tests pass locally.
• Lockfile Guard: If I added or modified a dependency in pyproject.toml, I have run uv lock --python 3.10 and committed the updated uv.lock file.
• I have added tests that prove my fix is effective or that my feature works.
• I have updated the documentation accordingly (if applicable).

📸 Screenshots / Visual Proof

The Production Autopilot report now provides a rich HTML dashboard comparing PaDiM, PatchCore, and EfficientAD, including localization and deployment metrics.

Recommended screenshots:

• production_autopilot_report.html
• EfficientAD anomaly heatmap
• EfficientAD localization mask/overlay
• PaDiM vs PatchCore vs EfficientAD comparison
• Autopilot model-selection summary