Settings
Model provenance and runtime configuration, read from the environment at boot.
Model
| File | cotton_mnv3s_fp16.tflite |
| Version | 930c20fa33fe |
| Size | 2.10 MB · 2.000 MiB |
| Format | tflite-fp16 |
| Runtime | ai_edge_litert |
| Input | 224 × 224 × 3 · float32 |
| Input range | 0–255 raw; the model rescales internally |
| Feature map | 7 × 7 × 576 |
Class mapping
0Aphids1Armyworm2Bacterial Blight3Healthy4Powdery Mildew5Target Spot
| Verified | yes |
| Method | probe_labels.py: sorted train dirs + empirical modal argmax over 360 images, agreed |
| Recorded | 2026-09-12 |
Verified against the dataset
Confirmed two ways by
tools/probe_labels.py — sorted class directories, and
the modal prediction over 60 images per class. Both agreed on all six indices.Runtime
| Grad-CAM | available · analytic (closed-form, no autodiff) |
| Max upload | 10 MB |
| Accepted | image/jpeg, image/png, image/webp |
| Interpreter threads | 1 |
| Low-confidence cutoff | 50% |
| Confidence precision | 1 decimal · FP16 noise ≈ 0.6% |
| App version | 0.1.0 |
Privacy
Images are classified in-process and are not sent anywhere — no external service, no network call at inference time. Nothing is written to disk in this phase. When persisted history arrives in Phase 6, uploads are re-encoded through Pillow on save, which strips EXIF metadata including GPS coordinates from phone photos.