Updates for January 31

[bfhealy]

General announcements

• Training has been re-running with adjusted hyperparameters after a collection of upgrades to resolve issues with the training and inference pipelines. These latest rounds of training has emphasized exposing each classifier to a larger amount of negative examples.

Pull requests

• Always upload batch classifications #226
• Download all available feature sets #227
• Additional check on existing sources' group membership #228
• Add argument to check 'labelled' box when uploading #229
• Get ZTF filters when starting with Fritz downloaded file #230
• Add option to delete existing classifications during upload #231
• Consistent/custom imputation strategy #207
• Update training pipeline #232
• Inference updates #234
• Add more subtlety to classification replacement #237

Issues

• Update feature imputation strategy #201
• Option to delete existing classifications when posting updated ones #212
• Include filter ID in feature catalogs #221
• Imputing features for inference sources #233
• Minimize imputation calls for training #235
• Minimize imputation calls for inference #236

Going forward

• The most recent training results are promising and are on track to yield a new active learning set by the end of the week. Concurrent with some additional hyperparameter adjustment, we will add code to combine predictions on multiple light curves for the same source before uploading the results to a new group on Fritz.