This repository contains sample training code and submissions for the 2025 HDR Scientific Mood (Modeling out of distribution) Challenge: Beetles as Sentinel Taxa. It is designed to give participants a reference for both working on the challenge, and also the expected publication of their submissions following the challenge (i.e., how to open-source your submission).
For your repository, you will want to complete the structure information below and add other files (e.g., training code):
submission
<model weights>
model.py
requirements.txt
We also recommend that you include a CITATION.cff for your work.
Note: If you have requirements not included in the whitelist, please check the issues on the challenge GitHub to see if someone else has requested it before making your own issue.
HDR-SMood-challenge-sample
│
├── baselines
│ ├── training
│ │ └── <MODEL NAME>
│ │ ├── evaluation.py
│ │ ├── model.py
│ │ ├── train.py
│ │ └── utils.py
│ └── submissions
│ └── <MODEL NAME>
│ ├── model.pth
│ ├── model.py
│ └── requirements.txt
├── .gitignore
├── LICENSE
└── README.md
To-Do: add notebook like last year (in a notebooks/ folder).
If you have uv simply run uv sync, otherwise you can use the requirements.txt file with either conda or pip.
An example training run can be executed by running the following:
python baselines/training/train.py
with uv do:
uv run python baselines/training/train.py
The training data is available on the Imageomics Hugging Face: Sentinel Beetles Dataset.
Aftering training, you can locally evaluate your model by running the following:
python baselines/training/evaluation.py
with uv do:
uv run python baselines/training/evaluation.py
Baselines built off of BioCLIPv2 and Dinov2:
Gu, Jianyang, et al. "Bioclip 2: Emergent properties from scaling hierarchical contrastive learning." arXiv preprint arXiv:2505.23883 (2025).
Oquab, Maxime, et al. "Dinov2: Learning robust visual features without supervision." arXiv preprint arXiv:2304.07193 (2023).
