Add default training pipeline #184
Conversation
| for class_nr in range(len(self.settings["classes"])) | ||
| ] | ||
|
|
||
| model = CatBoostClassifier( |
There was a problem hiding this comment.
should these model settings be configurable?
src/worldcereal/train/__init__.py
Outdated
There was a problem hiding this comment.
all this filtering of reference datasets and filtering of samples based on rules does not feel generic.
It's also very much tuned to Phase I extractions.
We should probably get rid of this at some point?
Maybe even have a separate repository where we train the global models?
There was a problem hiding this comment.
At some point, yes maybe, but this how the model is currently trained so we need to be transparent. I agree that these methods need to evolve in the coming months.
There was a problem hiding this comment.
main comment is whether we really want to commit all this Phase I related cleaning of datasets to this repository?
I guess in the end all default models will be trained based on Phase II RDM samples and extractions.
So perhaps training of global models should (for now) be done in a separate repository, where we import functionality from worldcereal-classification?
Just a suggestion...
i don't feel like setting up yet another repository, especially not right now. My thought was to add what there is now for transparency, but I can also accept to not merge it for the time being. |
This PR adds the training pipelines for the default cropland model trained on global Presto embeddings for the WorldCereal reference database.