🪸 Andromeida Reef Vision - ONNX Model Suite
This repository hosts optimized ONNX (Open Neural Network Exchange) weights for Andromeida Reef Vision Studio, enabling high-speed, ultra-lightweight desktop and edge inference via ONNX Runtime without requiring PyTorch or heavy CUDA drivers.
📦 Models Included
| File Name | Architecture | Task | Source Checkpoint | Size |
|---|---|---|---|---|
sam_image_encoder.onnx | Vision Transformer (ViT-B) | Deep Image Embedding extraction | reefsupport/CoralSCOP | 342 MB |
sam_mask_decoder.onnx | SAM Lightweight Mask Decoder | Instance Mask Generation & IoU Prediction | reefsupport/CoralSCOP | 16 MB |
bioclip_visual.onnx | ViT-B/16 Vision Transformer | Zero-Shot Coral Taxonomy Features | ReefNet/finetuned-bioclip | 329 MB |
coral_taxonomy_embeddings.npy | Pre-computed cosine text matrix | 18 Benthic Coral Taxa embeddings | Pre-encoded with BioCLIP | 36 KB |
bleaching_yolo11n.onnx | YOLO11n Classification | Coral Bleaching Condition Assessment | NMFS-OSI/yolo11n-cls-noaa-esd-coral-bleaching | 5.9 MB |
🙏 Credits & Author Attributions
We give full credit to the original research teams and institutions whose open-source foundation models power this suite:
1. CoralSCOP SAM ViT-B
- Model Card / Repo:
reefsupport/CoralSCOP - Developed by: Reef Support B.V. & EPFL Environmental Computational Science and Earth Observation Laboratory (ECEO).
- Base Architecture: Meta AI Segment Anything (
segment-anythingby Kirillov et al., ICCV 2023). - License: Apache 2.0 / Meta SAM Research License.
2. ReefNet Fine-Tuned BioCLIP
- Model Card / Repo:
ReefNet/finetuned-bioclip - Developed by: ReefNet Marine Vision Team.
- Base Architecture: BioCLIP ("BioCLIP: A Vision Foundation Model for the Tree of Life", Stevens et al., CVPR 2024 / Imageomics Institute).
- License: MIT License.
3. NOAA ESD Coral Bleaching Classifier
- Model Card / Repo:
NMFS-OSI/yolo11n-cls-noaa-esd-coral-bleaching - Developed by: National Oceanic and Atmospheric Administration (NOAA) Fisheries, Ecosystem Sciences Division (ESD).
- Base Architecture: Ultralytics YOLO11n Classifier.
- License: AGPL-3.0 / Open Government Work.
🚀 Quickstart with ONNX Runtime
import onnxruntime as ort
import numpy as np
from huggingface_hub import hf_hub_download
# Download and run Bleaching Classifier
model_path = hf_hub_download(
repo_id="harsh-awasthi/Andromeida-ReefVision-ONNX",
filename="bleaching_yolo11n.onnx"
)
session = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])
dummy_input = np.random.randn(1, 3, 224, 224).astype(np.float32)
output = session.run(None, {"images": dummy_input})
print("Bleaching probabilities (Healthy vs Bleached):", output[0])
🛡️ License
The ONNX weights provided here inherit the respective open-source licenses from their original creators. Please refer to the upstream repositories for commercial use and scientific citation guidelines.