This is a fundation model based solution for Solafune Finding Mining Sites. The pretrained model is based on the Prithvi-100M. Prithvi-100M is a temporal vision transformer model trained on US Harmonised Landsat Sentinel 2 (HLS) data, making it suitable for Sentinel 2 downstream tasks.
pip install -r requirements.txtDownload and unzip the training data from the competition page.
Rename the answer.csv to metadata.csv and put it in the training data folder.
Due to my git LFS quota, please download the pretrained model weight (Prithvi_100M.pt) from here and specify the path of Prithvi_100M.pt in setting.json.
This project uses wandb for logging. To log in, run the following command and follow the instructions.
wandb login- Automatic data preparation via huggingface datasets
- Stuctured codebase organized by PyTorch Lightning
- Pretrained model based on Prithvi-100M
- Training and evaluation visualization via wandb
- set the training data path in
setting.json - Run train.py
- Use inference.ipynb