| # Recording Ultralytics benchmark evidence |
|
|
| This package records each Ultralytics benchmark together with the runtime, |
| model, source revision, and Raspberry Pi host conditions needed to interpret |
| the result. |
|
|
| ## Usage |
|
|
| Run benchmark commands from the `yolo26-raspberry-pi-5/` runtime directory. |
| Reporting runs automatically through the Make target. See the |
| [Ultralytics benchmark documentation](../ultralytics/README.md) for benchmark |
| options. |
|
|
| By default, each invocation creates a UTC-timestamped directory under |
| `benchmarks/reporting/output/`. Set |
| `BENCHMARK_OUTPUT=/path/to/new-directory` to choose a specific new location. |
| The recorder never replaces an existing result. |
|
|
| ## Recorded data |
|
|
| Each evidence directory contains: |
|
|
| - `benchmark.log` - Combined Ultralytics standard output and standard error, |
| including accuracy metrics and stage timings. |
| - `results.json` - The command outcome, source revision, runtime identity, |
| complete model manifest, and host observations from before and after the |
| benchmark. |
|
|
| The host observations cover CPU policy, temperature, throttling, available |
| memory, filesystem cache, major page faults, and swap activity. |
|
|
| Every performance comparison must name its dataset and record the dataset |
| Hub revision, validation-configuration SHA-256, image count, and instance |
| count. The accepted baseline uses the pinned `coco-128-by-sa` revision selected |
| for publication. It is supplied explicitly during qualification and is not the |
| benchmark target's default dataset. The exact revision is recorded in |
| [`baseline.md`](baseline.md). |
|
|
| Update the accepted baseline only when review establishes a new result. |
|
|
| For a performance-sensitive pull request, summarize the reference and current |
| evidence with [`performance-report-template.md`](performance-report-template.md) |
| and attach both evidence directories to the pull request. |
|
|
