Small, on-device classifiers that identify wild berries, mushrooms, and medicinal plants from a photo — and refuse by default when unsure. Four tf_efficientnet_lite2 models (~9M params each, ~0.04B total) fine-tuned from timm/tf_efficientnet_lite2.in1k on iNaturalist research-grade observations. They run offline on a Raspberry Pi 5 + Hailo 8L in a handheld device; the same ONNX weights run on a free CPU.

🔗 Live demo: Forager's Field Station Space · Open dataset: HomesteaderLabs/forager-sightings

Architecture — single-expert routing (safety-first)

photo (224x224 RGB, ImageNet-normalized)
  -> domain_router_v2  (berry / mushroom / plant / other)
       conf < 0.74  or  "other"   -> ABSTAIN
       else route to the ONE expert that owns the domain:
         berry    -> berry_expert
         mushroom -> highvalue_expert
         plant    -> medicinals_expert
       expert below confidence gate -> ABSTAIN
  -> SAFE / CAUTION / DEADLY  (+ scientific name, look-alike, key difference)

Single-expert routing is a safety choice: an off-domain expert never gets to misclassify an input it doesn't own (the mushroom expert never sees a plant, so it can't call a poison hemlock "ramps"). All models emit raw logits; softmax, confidence gating, and abstention run downstream.

Models in this repo

Folder	Domain	Classes
`domain_router_v2/`	berry / mushroom / plant / other	4
`berry_expert/`	wild berries + toxic look-alikes	11
`highvalue_expert/`	chanterelle, morel, lion's mane, ginseng…	11
`medicinals_expert/`	wild medicinals + deadly look-alikes	21

Each folder has *_logits.onnx (CPU/edge), *_classes.json (label order), and *.pt (the fine-tuned timm checkpoint: model_state_dict + classes + arch). energy_thresholds.json holds per-expert in-domain energy percentiles for optional OOD gating.

Benchmarks

iNaturalist held-out validation: berry 92.1% · high-value 97.4% · medicinals 95.7% · router 92.9% (≥95% routing accuracy at conf 0.74). Toxic-as-edible false-accept rate: 0.00% on all three experts.

Real-world field photos: berry 95.5% · high-value 100% · medicinals 95.2%; across the held-out real set the stack never labelled a deadly specimen as edible.

Usage (ONNX, CPU)

python

import onnxruntime as ort, numpy as np, json
sess = ort.InferenceSession("berry_expert/berry_expert_logits.onnx",
                            providers=["CPUExecutionProvider"])
classes = json.load(open("berry_expert/berry_expert_classes.json"))
# x: float32 [1,3,224,224], ImageNet-normalized (Resize 255 -> CenterCrop 224)
logits = sess.run(None, {"input": x})[0][0]
probs = np.exp(logits - logits.max()); probs /= probs.sum()
print(classes[int(probs.argmax())], float(probs.max()))

⚠️ Safety

Identification aid only — never an authority. Wild mushroom and plant ID carries fatal risk. No output should be acted on — including any consumption decision — without independent verification by a qualified expert. Amatoxin poisoning (Amanita, Galerina, Conocybe) is lethal with no reliable field antidote. The maintainers accept no liability for decisions made from model output.

License & citation

Apache-2.0. Training data from iNaturalist observers (CC-BY / CC0); weights are an original derivative released independently under Apache-2.0.

HomesteaderLabs (2026). Forager's Field Station Field ID Models. https://homesteaderlabs.com

HomesteaderLabs/forager-field-station-models

Forager's Field Station — Field ID Models