Classifier not updating

[Chelsea Brown]

I have been waiting several days for the classifier to update so I can use the annotations I have entered (34 images with 100 points = 3400 points) to auto-annotate my remaining images. 

It reads that it was trained on 22 images and has 68% accuracy, dated Sept 2, 2025. Has not updated since despite me completing a total of 34 images. 

Why has it not updated to incorporate the more recent annotations so that I can start using it? This is making the site pretty unusable as a trained annotator. 

[Chelsea Brown]

Reaching back out about this. 

[Stephen Chan]

Hi Chelsea - apologies for the wait; I can be a bit slow to respond to questions.

I'll need your source's ID-number or name to be sure of this situation, but one thing that can commonly confuse folks is the difference between "Last classifier saved" and "Last classifier trained". CoralNet should be consistent about training a new classifier when your confirmed-image count has gone up enough, but that newly-trained classifier may not get saved if the evaluated accuracy hasn't improved enough since the previous classifier.

You can get more info on recent classifier trainings by clicking the "Jobs" tab at the top of your source, and applying the filter "Type" = "Train classifier".

[Chelsea Brown]

Source Name: Quad Photos

I have included some of things you mentioned as I can see them. It has remained at 68% from photo 22 to photo 34 (completed on Sept 4). 

• Last classifier saved: Sept. 2, 2025, 12:34 p.m.
• Last classifier trained: Sept. 2, 2025, 1:40 p.m.
• Feature extractor: EfficientNet (default)
• Confidence threshold: 100%

[Stephen Chan]

OK - here's what the Jobs tab for this source shows (screenshot attached). After the classifier with 22 images was saved, there was another training about an hour later with a higher number of images (maybe close to 34) where the accuracy metric ended up being about the same as the previous classifier, 68%, so the classifier resulting from that last training was not saved.

If you can get to 38 or more images confirmed, a new training should run, and that classifier's accuracy might (or might not) end up high enough to save this time.