verdant-plant-model

Verdant — on-device plant model bundle

Two ONNX classification heads used by the Verdant mobile plant-diary app. All inference runs locally on the phone; user photos never leave the device. The app downloads this bundle once on first launch and caches it.

config.json carries the full label maps, preprocessing parameters (resize/crop/channel order/normalization) and per-head confidence thresholds. version.json carries sizes and SHA-256 checksums for resumable, verified downloads.

Quantization notes

• Species (MobileViT → int8): dynamic int8 quantization is a clean win — 3.6× smaller with top-1 parity vs fp32 (40/40 agreement on random inputs). Transformer matmuls quantize well.
• Leaf condition (MobileNetV2 → fp32, kept): dynamic int8 quantization of MobileNetV2's depthwise convolutions collapsed accuracy (0/60 top-1 agreement) and ran slower on CPU, so this head ships fp32 — consistent with the project spec (fp32 default, quantization optional).

Preprocessing (from config.json)

Species: shortest-edge resize 288 → center-crop 256 → x/255 → BGR channel order, no mean/std. Leaf: shortest-edge resize 256 → center-crop 224 → x/255 → RGB → normalize mean/std = 0.5.

Output is raw logits; apply softmax. If top-1 probability is below the head's confidence_threshold, show candidate alternatives or an "unsure" state and prompt the user to retake the photo.

Provenance & limitations

MVP heads are derived from public checkpoints: dybdyb/mobilevit-houseplants (species) and linkanjarad/mobilenet_v2_1.0_224-plant-disease-identification (leaf condition, PlantVillage-trained).

The leaf-condition head is trained on lab-style crop images. Per the project spec, it must be revalidated and retrained on field data (PlantDoc + own collection) before production — lab accuracy does not reflect real-world phone photos. Species coverage is a starter set of popular houseplants and grows iteratively.