Pre-trained model weights / pre-computed predictions / standalone training & inference script

[legel]

Awesome project.

Would love to be able to actually work with the model, right now that indeed appears impossible as suggested by #6.

Specific requests:

1. Pre-trained models weights from your best run, downloadable from HF or elsewhere
2. Pre-computed model predictions (e.g. .tif files) across CONUS, and/or
3. Standalone training / inference scripts not requiring Beaker, with documentation

The above is probably not that hard, but right now it's totally unclear how to move forward with this, even if just for evaluation of your work.

[APatrickJ]

Thank you for the kind words and for making those requests clear.

On (1), we can work to get the model weights available on HF for your evaluation.

As for precomputed predictions for (2), are there any specific locations and times that you'd like to see first? All of CONUS at 10m resolution will currently be quite expensive for us to produce, but in the near term we can try to do cover specific spatiotemporal windows for smaller AOIs.

As for standalone scripts (3), we can make something available that doesn't require Beaker. Would you be comfortable taking Python code and running it on hardware / the cloud provider of your choice?