| # Local deployment |
|
|
| ## Desktop / local service integration |
| The supported entry point is `src.inference.predict_wav`. Install requirements and keep `model.h5`, `config.json`, `labels.json`, and `src/` together. There is no network dependency in the inference code. |
|
|
| ## Android / Kernel integration notes |
| Kernel's supplied architecture requires all live inference to remain on-device. This package is a Python/PyTorch training artifact, not yet a drop-in Android artifact. Before shipping: |
| 1. Freeze the audio frontend and compare it against the training implementation using a golden WAV fixture. |
| 2. Convert the model to an Android-supported runtime (e.g. LiteRT/TFLite) and test numerical parity; do not rename `.pt`/`.h5` to `.tflite`. |
| 3. Benchmark on the actual iQOO 15: latency, thermal/battery, noise, sample-rate path and memory. |
| 4. Validate with field-collected grain-pest recordings and verified physical inspections. |
| 5. Keep outputs experimental until false-positive, false-negative, uncertainty and cross-device performance are evaluated. |
| 6. Store model version, audio quality and human-verified outcome separately. Never treat a Clean prediction as proof of pest-free grain. |
|
|
| ## Suggested production architecture |
| Microphone -> signal-quality gate -> exactly matched local preprocessing -> local model -> experimental result UI -> local SQLite record. GCP services, if enabled, should sync only authorized metadata and must not receive raw audio or perform inference. |
|
|
