Skip to content

feat(hailo): add PatchCore Hailo export on latest main - #169

Closed
DeepKnowledge1 wants to merge 21 commits into
mainfrom
feat/hailo-patchcore-main-clean
Closed

feat(hailo): add PatchCore Hailo export on latest main#169
DeepKnowledge1 wants to merge 21 commits into
mainfrom
feat/hailo-patchcore-main-clean

Conversation

@DeepKnowledge1

@DeepKnowledge1 DeepKnowledge1 commented Aug 28, 2026

Copy link
Copy Markdown
Owner

Clean Hailo PatchCore integration based directly on the latest main.

This intentionally avoids carrying the large shared-file changes from feat/hailo-patchcore-e2e into detect.py, train.py, config.py, and autopilot.py.

Preserved from main:

  • PaDiM
  • PatchCore
  • EfficientAD
  • Kria/KV260
  • existing generic inference pipeline and Hailo backend

Added from the working Hailo branch:

  • Hailo-specific PatchCore export graph
  • Hailo HEF exporter integration for PatchCore
  • Hailo PatchCore compilation script

The original working branch feat/hailo-patchcore-e2e remains untouched as the backup.

Hardware validation is still required on a Hailo host; this PR intentionally minimizes shared-code regression risk first.

Copy link
Copy Markdown
Owner Author

Updated the clean branch with hardware-free Hailo validation and fixed calibration preprocessing. Hailo calibration tensors now use the same ImageNet normalization contract as AnomaVision inference. Added scripts/validate_hailo_vs_onnx.py for HEF/ONNX shape/output contract checks and ONNX reference inference; it explicitly does not claim numerical HEF equivalence without HailoRT/device execution. Added tests verifying that already-preprocessed NCHW tensors are only transposed for Hailo and that raw images are normalized exactly once. The branch remains based directly on main and does not include the old detect.py/train.py/config.py/autopilot.py changes.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant