| # Detect-Person miner (element `manak0/Detect-Person`) |
|
|
| Public-track TurboVision miner: single-class person detection. |
| Base model: YOLO11n (COCO-pretrained, ~5.6 MB — well under the 30 MB element cap), |
| filtered to the person class, emitted with `cls_id=0`. |
|
|
| Element constraints (from the live manifest): |
|
|
| - pillars: map50 x0.6 + false_positive x0.4 (false_positive = 1 - FPs-per-image/10) |
| - baseline theta to beat: ~0.368 |
| - p95 latency <= 100 ms/frame on 2 vCPU (automated compliance loop) |
| - max model size 30 MB; validator preproc: 5 fps, long side 1280, rgb-01 |
|
|
| ## Files |
|
|
| - `miner.py` — required entrypoint (`Miner.predict_batch`), prefers `person.onnx` over `yolo11n.pt` |
| - `chute_config.yml` — Chutes image/runtime config |
| - `yolo11n.pt` — base weights (replace with fine-tuned weights to beat other miners) |
| - `export_onnx.py` — exports `person.onnx` for faster CPU inference |
| - `benchmark_latency.py` — local replica of the 2 vCPU / 100 ms p95 latency gate |
|
|
| ## Workflow |
|
|
| ```bash |
| # from the repo root, with the uv venv active (uv sync already done) |
| cd my_miners/detect_person |
| |
| # 1. ONNX export (required for CPU speed; .pt alone fails the latency gate) |
| python export_onnx.py --imgsz 480 |
| |
| # 2. latency gate check, pinned to 2 CPUs like the compliance loop |
| taskset -c 0,1 python benchmark_latency.py --frames 100 --imgsz 480 |
| |
| # 3. deploy: upload to HF, build chute, warm + health-check, commit on-chain |
| cd ../.. |
| sv -v deploy-os-miner --model-path my_miners/detect_person --element-id manak0/Detect-Person |
| |
| # dry-run variants |
| sv -v deploy-os-miner --model-path my_miners/detect_person --element-id manak0/Detect-Person --no-commit |
| ``` |
|
|
| Required `.env` (see repo README.md): `BITTENSOR_WALLET_COLD`, `BITTENSOR_WALLET_HOT`, |
| `CHUTES_API_KEY`, `HF_USER`, `HF_TOKEN`, `CHUTES_HF_TOKEN` (read-only fine-grained token), |
| `SCOREVISION_NETUID=44`. Hotkey must be registered on netuid 44. |
|
|
| ## Measured latency (this machine, taskset 2 CPUs, 60 frames) |
|
|
| | Model | imgsz | p50 | p95 | Gate (<=100 ms p95) | |
| |---|---|---|---|---| |
| | yolo11n.pt (torch CPU) | 640 | 131 ms | 150 ms | FAIL | |
| | person.onnx | 640 | 98 ms | 145 ms | FAIL | |
| | person.onnx | 512 | 57 ms | 92 ms | pass (thin margin) | |
| | person.onnx | 480 | 52 ms | 78 ms | PASS (default) | |
|
|
| The compliance CPU model is unknown, so the 480 export is the default for |
| headroom. Re-export at 512 only if you accept the thinner margin. |
|
|
| ## Tuning notes |
|
|
| - `SV_CONF_THRESHOLD` (default 0.35): raise to cut false positives (each FP/image |
| costs 0.1 on the 0.4-weighted FP pillar), lower to lift recall/mAP. |
| - The dataset is synthetic (SAM3-annotated, 302 images). COCO-pretrained YOLO11n |
| should clear the ~0.368 baseline; to win the element, fine-tune on person data |
| matching the synthetic distribution and drop the weights in as `yolo11n.pt` |
| (or re-export `person.onnx`). |
| - Commit early: on-chain commit block is the tie-breaker under model-copy protection. |
|
|
