EarthPT

A simple repository for training time series large observation models. This repository began its life as Andrej Karpathy's nanoGPT, and has been altered so that it is usable for time series data. train.py reproduces EarthPT-700M when trained on 14B time series 'tokens' of ClearSky EO data within the TL UK National Grid tile. When run, train.py takes ~5 days to achieve Chinchilla 🐭 completion on a single 8xA100 40GB node. Within train.py you will find a ~300-line boilerplate training loop and within model.py you will find a ~300-line GPT model definition with an MLP tokeniser and a regressive loss.

We have purposefully kept the code as simple and hackable as possible so that it is easy for all to hack this base to their needs. We release this code under the MIT licence in the hope that it will prove useful to others working on EO and timeseries large observation models.

install

Dependencies:

• pip install -r requirements.txt

results

Our EarthPT-700M model is able to predict future satellite passes well into the future, and also learns semantically meaningful information about the timeseries that it is fed:

You can find a plot with less angular momentum and further results in our paper here.

pretrained weights

You can find our weights for all the EarthPT models on HuggingFace 🤗.

citation

If you find EarthPT useful in your work please do drop us a cite:

@article{ref_smith2023,
    author = {Smith, M. J. and Fleming, L. and Geach, J. E.},
    title = {{EarthPT: a time series foundation model for Earth Observation}},
    journal = {arXiv},
    year = {2023},
    eprint = {2309.07207},
    doi = {10.48550/arXiv.2309.07207}
}

This work is also in the proceedings of the 2023 CCAI NeurIPS workshop.