Hybrid Spatial Representations for Species Distribution Modeling

We use the same version of iNaturalist data as in SINR, which can be obtained as follows:

curl -L https://data.caltech.edu/records/b0wyb-tat89/files/data.zip --output data.zip
unzip -q data.zip
rm data.zip

Afterwards, install requirements with:

conda create --name hsrsdm python=3.10
conda activate hsrsdm
pip install -r requirements.txt
pip install torch==2.1.2+cu118 torchvision==0.16.2+cu118 --extra-index-url https://download.pytorch.org/whl/cu118
conda install -c "nvidia/label/cuda-11.8.0" cuda-toolkit
pip install ninja git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch

Then run python train_and_evaluate_models.py to get results. Key hyper-parameters (learning rate, implicitness, and observation cap) can be directly adjusted in the file.

Model Zoo

The model zoo can be found at Open Science Framework: https://osf.io/pbk9a/?view_only=e4b3ced52eeb460a9f8cd8dd1569a0df

Acknowledgements

This codebase borrows from SINR, NeRFStudio, and tiny-cuda-nn. We would like to thank the authors of those works.