TsinghuaCorals/bioclip-2.5-reefnet-bleaching-lora

BioCLIP 2.5 ViT-H — ReefNet Coral Bleaching LoRA (binary)

A binary healthy hard coral vs bleached hard coral classifier built on top of the ReefNet species LoRA model `BobDerBaum/bioclip-2.5-vith14-reefnet-lora`, which provides the fine-tuned vision-encoder LoRA adapters. Only a small linear head is trained on top (frozen backbone + LoRA + 2-way linear classifier).

Artifacts

• —`bleaching_bioclip_lora_best.pt` — a portable Lightning checkpoint containing the vision LoRA adapters + the binary classifier only (~94 MB). The frozen BioCLIP 2.5 base is pulled from HuggingFace at load time. Load with:

  import torch
  from species_atlas.bleaching.model import BackboneLinearProbeModel

  model = BackboneLinearProbeModel.load_from_checkpoint(
      "BobDerBaum/bioclip-2.5-reefnet-bleaching-lora/bleaching_bioclip_lora_best.pt",
      weights_only=False,
  ).eval()
  # or download via huggingface_hub.hf_hub_download first

  with torch.no_grad():
      logits = model(images)          # (N, 2): [healthy, bleached]

Label order: ["healthy hard coral", "bleached hard coral"].

Training

• —Backbone: frozen imageomics/bioclip-2.5-vith14 + vision LoRA (rank 32, alpha 64) loaded from the ReefNet species LoRA run (62yuzh9j).
• —Head: linear classifier on the 1024-d vision features.
• —Data: NOAA-PIFSC-ESD Coral Bleaching dataset (binary). AdamW lr 1e-3.
• —3 epochs (early stopped at epoch 6 of a longer schedule).

Metrics (run nxhcu8k7)

Best of the three bleaching backbones tried (frozen BioCLIP 0.8454; DINOv3-L 0.8441; this LoRA model 0.9016 test accuracy). The species-level backbone LoRA transfers well to the bleaching binary task.

Data

Training images: the NOAA-PIFSC-ESD bleaching dataset. Species backbone training data: `BobDerBaum/reefnet_species_images`.