Perturb / remove latlons

WorldCereal / presto-worldcereal

4 stars 0 forks source link

Perturb / remove latlons #48

Open gabrieltseng opened 7 months ago

gabrieltseng commented 7 months ago

Might be making the model prone to overfitting for poorly distributed labels

kvantricht commented 4 months ago

@gabrieltseng is this something that we should actively look into? (cc @cbutsko )

gabrieltseng commented 4 months ago

It seems like spatial overfitting is not a major problem - I think this is still an interesting experiment to run, but lower priority

kvantricht commented 4 months ago

Would that conclusion also hold for crop type which currently has much more spatial bias? With the crop/no-crop data being fairly well spread geographically, I guess it's not a major surprise that spatial overfitting is not really a problem. I'm still wondering though if the same will be true for crop type.

gabrieltseng commented 4 months ago

Good point - no, I am not sure it would.

Perhaps we should keep this issue open until we have a better sense of spatial generalization for the crop type task.

kvantricht commented 3 months ago

@cbutsko please report your findings here on some of the work with fixed lat/lon tests you're doing.

gabrieltseng commented 3 months ago

[ ] For removing latlons, @gabrieltseng to make the comment in https://github.com/WorldCereal/presto-worldcereal/pull/71/files#r1655642661 programmatic.