Access requires agreeing to the upstream terms
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This is a Core ML conversion of a GATED upstream model whose training data carries GBIF and iNaturalist terms. By requesting access you confirm you have reviewed the upstream model's terms and the source data licences, and that your use complies with them. Redistribution of these weights is your responsibility, not the converter's.
Log in or Sign Up to review the conditions and access this model content.
Onyx Plant ID 14k — Core ML
On-device plant species identification: 14,829 species, every label a Latin binomial. ViT-Base, ~190 MB compiled at fp16, fully offline through Core ML / Vision.
Licensing — read this before using or redistributing
This repository contains no original model weights. It is a format conversion of
juppy44/plant-identification-2m-vit-b,
and the licence of that model governs these files.
Three distinct layers apply, and they are not the same licence:
- The upstream model is published as Apache-2.0 and is access-gated. Gating is the author's distribution choice; converting the weights does not transfer permission to bypass it. This repository is therefore gated too — a deliberate mirror of the upstream posture, not an inconvenience.
- The training data (~2M images) carries GBIF and iNaturalist terms. Those terms attach to data provenance and are not superseded by the model's Apache header. Many iNaturalist observations are CC-BY or CC-BY-NC; some are all-rights-reserved.
- This conversion (the Core ML graph, the preprocessing folded into it, the build script) is offered under Apache-2.0 to the extent it is separable from the above.
If you need weights with a clean, single-licence provenance, use
wabibito/Onyx-PlantNet300K-CoreML
instead: 1,081 species, Apache-2.0, ungated, PlantNet-300K (Garcin et al., NeurIPS 2021).
Attribution
- Model: juppy44,
plant-identification-2m-vit-b. - Data: GBIF / iNaturalist contributors.
- Conversion: Onyx.
Conversion notes
- Softmax is in the graph exactly once. The HF model returns logits inside a dataclass; the wrapper unwraps to a plain tensor and softmax is applied after. Applying it in both places flattens the distribution — softmax of a softmax — so every prediction reads as low confidence.
- Label order is numeric, not lexical.
id2labelis keyed by string indices; sorting them as strings puts "10" before "2" and silently scrambles all 14,829 classes. - ImageNet normalisation is folded into the Core ML input.
Verified on the compiled model: 14,829 classes, probability mass 0.999.
Honest limits
- No accuracy metrics. The upstream card publishes none ("no clean evaluation split"), so none is claimed here.
- 14,829 species is broad but a fraction of ~400,000 known plants.
- Confidence is spread over 14,829 classes, so a low top-1 here is not comparable to the same number from a 1,000-class model.
Never treat a photo-based identification as authority on edibility, medicinal use or safety. Look-alikes exist; confirm with a local expert or field guide.
Model tree for wabibito/Onyx-PlantID-14k-CoreML
Base model
google/vit-base-patch16-224