| # YOLO26n model converter |
|
|
| The converter turns a compatible YOLO26n PyTorch checkpoint into the |
| `model.bin` and `manifest.json` files consumed by the native runtime. It uses |
| the Python 3.13 standard library and does not import Torch or Ultralytics. |
|
|
| Export and validate a checkpoint from the `yolo26-raspberry-pi-5/` runtime |
| directory: |
|
|
| ```sh |
| make model RESOLUTION=640 CHECKPOINT=/absolute/yolo26n.pt |
| make model-check RESOLUTION=640 CHECKPOINT=/absolute/yolo26n.pt |
| ``` |
|
|
| Supported resolutions are 320, 480, 640, 800, 960, and 1280. Each converted |
| model has one static input resolution. The same native runtime library can |
| execute models at any supported resolution. |
|
|
| ## Conversion pipeline |
|
|
| 1. `checkpoint_decoder` reads the supported Torch archive without importing |
| executable model code. |
| 2. `model_validation` validates the model type, operator topology, routing, |
| tensor inventory, class labels, and detection-head geometry required by the |
| fixed runtime. |
| 3. `weight_lowering` folds BatchNorm, converts convolution layouts, and |
| creates the split projections used by the runtime. |
| 4. `model_format` serializes the named weight slots, resolution, and class |
| metadata to `model.bin`. |
| 5. `export_yolo26n` publishes `model.bin` and `manifest.json` together. |
|
|
