vincent-espitalier/dino-v2-reg4-with-plantclef2024-weights

This repo contains pre-trained weights for the "reg4" version of the DINOv2 model converted into a format that can be used by candle.

The model was finetuned on about 1.4 million images covering 7806 vascular species to identify plants, and proposed to the participants of the PlantCLEF 2024 competition. See the HuggingFace's page of the PlantCLEF 2024 competition. The original PyTorch weights can be found on Zenodo. The model tag is vit_base_patch14_reg4_dinov2_lvd142m_pc24_onlyclassifier_then_all

The extraction script used the code from pytorch-to-safetensor-converter

Citing DINOv2 reg4 finetuned on PlantCLEF 2024 dataset

@misc{goeau_2024_10848263,
  author       = {Goëau, Hervé and
                  Lombardo, Jean-Chirstophe and
                  Affouard, Antoine and
                  Espitalier, Vincent and
                  Bonnet, Pierre and
                  Joly, Alexis},
  title        = {{PlantCLEF 2024 pretrained models on the flora of 
                   the south western Europe based on a subset of
                   Pl@ntNet collaborative images and a ViT base patch
                   14 dinoV2}},
  month        = mar,
  year         = 2024,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.10848263},
  url          = {https://doi.org/10.5281/zenodo.10848263}
}