Generate model weights locally

Remove the converted model binary from the repository. Download and verify the official YOLO26n checkpoint, then generate the benchmark artifact locally. Keep the AI Portal accuracy report aligned with the current coco-128-by-sa measurement.

README.md

@@ -13,7 +13,7 @@ The runtime, converter, Python API, and Ultralytics adapter correspond to
 commit `9056a527deaf2adc6123f43424d07672753a9524` of
 [Arm-Runtime-for-YOLO26-on-Raspberry-Pi-5](https://github.com/arm/Arm-Runtime-for-YOLO26-on-Raspberry-Pi-5).
 
-## Benchmark the supplied converted model
+## Benchmark the standard model
 
 Internet access is required during setup and while downloading the official
 YOLO26n checkpoint and the validation dataset. Authenticate with `hf auth
@@ -26,9 +26,10 @@ python3 repro.py benchmark
 
 `repro.py setup` installs Ultralytics 8.4.108 and applies the included patch
 that registers this runtime as an Ultralytics export format and inference
-backend. The benchmark therefore uses Ultralytics `benchmark()`, with an
-explicit integration patch because upstream Ultralytics does not natively know
-how to load this runtime.
+backend. On its first run, the benchmark downloads the verified official
+YOLO26n checkpoint and converts it locally under `build/`. It then uses
+Ultralytics `benchmark()`, with an explicit integration patch because upstream
+Ultralytics does not natively know how to load this runtime.
 
 Before starting, the benchmark checks for filesystem and memory conditions
 that could distort performance, including RAM-backed model or dataset paths,
@@ -55,9 +56,10 @@ python3 repro.py benchmark
 
 Export creates a runnable model directory under
 `build/arm_yolo26n_pi5_model/`. Benchmark automatically selects it and checks
-that the source checkpoint matches the converted model. No PyTorch checkpoint
-is distributed in this repository; the stock benchmark downloads and verifies
-the official checkpoint when needed.
+that the source checkpoint matches the converted model. Neither a PyTorch
+checkpoint nor converted model weights are distributed in this repository;
+the stock benchmark downloads, verifies, and converts the official checkpoint
+when needed.
 
 ## About this version
 

metadata.yaml

@@ -2,10 +2,10 @@ schema_version: 1.0.0
 report_type: vision-detection
 task_type: object-detection
 title: YOLO26n for Raspberry Pi 5
-description: An Arm-optimized YOLO26n object-detection model for CPU inference on Raspberry Pi 5 using the project-specific Arm Runtime.
+description: An Arm-optimized YOLO26n object-detection model generated locally for CPU inference on Raspberry Pi 5 using the project-specific Arm Runtime.
 family_name: YOLO26
 id: Arm/Arm-Runtime-for-YOLO26-on-Raspberry-Pi-5
-filename: runtime/arm_yolo26n_pi5_model/model/model.bin
+filename: build/arm_yolo26n_pi5_model/model/model.bin
 base_model_id: yolo26n
 vendor: Ultralytics
 base_model_url: https://docs.ultralytics.com/models/yolo26/

model/README.md

@@ -4,3 +4,4 @@ To export another compatible YOLO26n model, place its checkpoint here as
 `yolo26n.pt`, then run `python3 repro.py export` from the repository root.
 
 No PyTorch checkpoint is distributed in this repository.
+Converted model weights are also generated locally and are not distributed.

repro.py

@@ -30,7 +30,7 @@ STOCK_CHECKPOINT_SHA256 = (
     "9b09cc8bf347f0fc8a5f7657480587f25db09b34bf33b0652110fb03a8ad4fef"
 )
 STOCK_CHECKPOINT = ROOT / "downloads/yolo26n.pt"
-SUPPLIED_MODEL = ROOT / "runtime/arm_yolo26n_pi5_model"
+RUNTIME_TEMPLATE = ROOT / "runtime/arm_yolo26n_pi5_model"
 EXPORTED_MODEL = ROOT / "build/arm_yolo26n_pi5_model"
 
 
@@ -144,9 +144,9 @@ def ensure_stock_checkpoint() -> Path:
     return STOCK_CHECKPOINT
 
 
-def export_model() -> None:
+def export_model(checkpoint: Path | None = None) -> None:
     python = VENV / "bin/python"
-    checkpoint = ROOT / "model/yolo26n.pt"
+    checkpoint = checkpoint or ROOT / "model/yolo26n.pt"
     if not python.is_file():
         raise SystemExit("run 'python3 repro.py setup' first")
     if not checkpoint.is_file():
@@ -176,7 +176,7 @@ def export_model() -> None:
             env=environment,
         )
         assembled = temporary / EXPORTED_MODEL.name
-        shutil.copytree(SUPPLIED_MODEL, assembled)
+        shutil.copytree(RUNTIME_TEMPLATE, assembled)
         shutil.rmtree(assembled / "model")
         shutil.copytree(converted_model, assembled / "model")
         assembled.rename(EXPORTED_MODEL)
@@ -293,10 +293,13 @@ def benchmark() -> None:
         raise SystemExit("run 'python3 repro.py setup' first")
     dataset_yaml = ensure_dataset(python)
 
-    model = EXPORTED_MODEL if EXPORTED_MODEL.is_dir() else SUPPLIED_MODEL
-    checkpoint = ROOT / "model/yolo26n.pt" if model == EXPORTED_MODEL else ensure_stock_checkpoint()
-    if not checkpoint.is_file():
-        raise SystemExit("the exported model requires its source checkpoint at model/yolo26n.pt")
+    custom_checkpoint = ROOT / "model/yolo26n.pt"
+    if EXPORTED_MODEL.is_dir():
+        checkpoint = custom_checkpoint if custom_checkpoint.is_file() else ensure_stock_checkpoint()
+    else:
+        checkpoint = ensure_stock_checkpoint()
+        export_model(checkpoint)
+    model = EXPORTED_MODEL
     preflight((model.resolve(), DATASET_DIR.resolve()))
 
     governor = Path("/sys/devices/system/cpu/cpu0/cpufreq/scaling_governor")

results/REFERENCE.md

@@ -21,6 +21,6 @@ writeback data, no refaults or reclaim scans, and no swap activity. The model
 and dataset were on ext4. The CPU governor was `ondemand`, and the device
 reported no throttling (`0x0`).
 
-The export path was also verified with the same stock checkpoint. It reproduced
-the packaged `model.bin` byte-for-byte and completed the benchmark with the same
-0.4866 mAP50-95 result.
+The local export path was verified with the same stock checkpoint and completed
+the benchmark with the same 0.4866 mAP50-95 result. Converted model weights are
+not distributed in this repository.

runtime/arm_yolo26n_pi5_model/model/model.bin

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:031991352d4c1bcca4ad5b6f7023b6e3d66bdf896b476c66b851b1a29c5bdaff
-size 9640064