eblume/blumeops
1
1
Fork 0
Code Issues Pull requests Projects Releases 83 Packages Wiki Activity Actions

Add Apple Silicon ZMQ detector for Frigate #206

Merged
eblume merged 7 commits from feature/frigate-zmq-detector into main 2026-02-17 19:03:29 -08:00
Conversation 0 Commits 7 Files changed 8 +137 -14
eblume commented 2026-02-17 17:12:38 -08:00
Owner
Copy link

Summary

  • New frigate_detector ansible role deploys the apple-silicon-detector as a LaunchAgent on indri
  • Switches Frigate from ONNX CPU detector (~117ms) to ZMQ detector backed by CoreML/Neural Engine (~15ms)
  • Removes detect FPS cap (no longer needed with fast inference)
  • Updates Frigate docs and adds changelog fragment

Deployment

Phase 1: Deploy detector on indri (one-time setup + ansible)

ssh indri 'git clone https://github.com/frigate-nvr/apple-silicon-detector.git ~/code/3rd/apple-silicon-detector'
ssh indri 'cd ~/code/3rd/apple-silicon-detector && make install'
mise run provision-indri -- --tags frigate_detector --check --diff  # dry run
mise run provision-indri -- --tags frigate_detector                 # apply
ssh indri 'launchctl list mcquack.eblume.frigate-detector'          # verify running
ssh indri 'tail ~/Library/Logs/mcquack.frigate-detector.out.log'    # verify bound

Phase 2: Test connectivity

kubectl --context=minikube-indri -n frigate exec deploy/frigate -- nc -vz host.minikube.internal 5555

Phase 3: Deploy Frigate config (branch workflow)

argocd app set frigate --revision feature/frigate-zmq-detector && argocd app sync frigate

Phase 4: Post-deploy checks

  • Pod starts, no config errors
  • /api/stats shows detector type zmq, inference_speed ~15ms
  • detect_fps uncapped
  • Recordings and MQTT events flowing
  • After merge: argocd app set frigate --revision main && argocd app sync frigate

🤖 Generated with Claude Code

## Summary - New `frigate_detector` ansible role deploys the [apple-silicon-detector](https://github.com/frigate-nvr/apple-silicon-detector) as a LaunchAgent on indri - Switches Frigate from ONNX CPU detector (~117ms) to ZMQ detector backed by CoreML/Neural Engine (~15ms) - Removes detect FPS cap (no longer needed with fast inference) - Updates Frigate docs and adds changelog fragment ## Deployment ### Phase 1: Deploy detector on indri (one-time setup + ansible) ```fish ssh indri 'git clone https://github.com/frigate-nvr/apple-silicon-detector.git ~/code/3rd/apple-silicon-detector' ssh indri 'cd ~/code/3rd/apple-silicon-detector && make install' mise run provision-indri -- --tags frigate_detector --check --diff # dry run mise run provision-indri -- --tags frigate_detector # apply ssh indri 'launchctl list mcquack.eblume.frigate-detector' # verify running ssh indri 'tail ~/Library/Logs/mcquack.frigate-detector.out.log' # verify bound ``` ### Phase 2: Test connectivity ```fish kubectl --context=minikube-indri -n frigate exec deploy/frigate -- nc -vz host.minikube.internal 5555 ``` ### Phase 3: Deploy Frigate config (branch workflow) ```fish argocd app set frigate --revision feature/frigate-zmq-detector && argocd app sync frigate ``` ### Phase 4: Post-deploy checks - [ ] Pod starts, no config errors - [ ] `/api/stats` shows detector type zmq, inference_speed ~15ms - [ ] detect_fps uncapped - [ ] Recordings and MQTT events flowing - [ ] After merge: `argocd app set frigate --revision main && argocd app sync frigate` 🤖 Generated with [Claude Code](https://claude.com/claude-code)
Moves object detection from ONNX CPU (~117ms/frame) to the
apple-silicon-detector running natively on indri via CoreML/Neural Engine
(~15ms), communicating with Frigate over ZMQ (tcp://host.minikube.internal:5555).

- New ansible role `frigate_detector` with LaunchAgent
- Switch Frigate configmap from ONNX to ZMQ detector
- Remove detect FPS cap (no longer needed with fast inference)
- Update docs and add changelog fragment

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Runs the detector script directly via `uv run --with` with inline
dependencies, eliminating the need for `make install` and a managed venv.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
YOLO-NAS has dynamic output shapes incompatible with CoreML EP, and the
apple-silicon-detector has no yolonas post-processor. YOLOv9-t works with
CoreML and uses the supported yolo-generic model type.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
CoreML runs 468/662 model nodes with the rest on CPU, yielding
~50-80ms inference rather than the initially estimated ~15ms.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Add driveway zone from UI config with review detection filtering
- Upgrade model from YOLOv9-t (7.7M params) to YOLOv9-m (51M params)
  for improved nighttime/low-light detection capability
- YOLOv9-t produced max 5.7% confidence on nighttime IR footage

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
eblume merged commit 5f9b024b4a into main 2026-02-17 19:03:29 -08:00
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
eblume/blumeops!206
No description provided.